Keywords: Soft Robot Materials and Design, Mechanism Design, Biologically-Inspired Robots
Abstract: Compared to traditional robots based on rigid links, soft robots have the advantage of high flexibility and high adaptability to complex environments. However, due to their high compliance, it is difficult to model and design soft robotic structures using the classic rigid-body mechanism theory. To cope with this problem, we utilize the continuum-structure-based topology optimization methods in our research to improve the design efficiency of soft robots. As prerequisite for the optimization algorithm, a geometrically nonlinear finite-element-modeling (FEM) tool has been implemented in a unified design framework in MATLAB to realize high-fidelity mechanics modeling of soft robots. By introducing additional artificial springs in the compliant-mechanism-based topology optimization problem, we can successfully achieve highly flexible soft robotic structures with balanced stress distribution. A design case of soft robotic finger is presented in this extended abstract to demonstrate the performance of our optimization algorithm. In addition, we are also developing multi-objective topology optimization methods in our current work to realize multi-functional soft robotic structures.

Keywords: Failure Detection and Recovery, Manipulation Planning, Learning from Experience
Abstract: This project explores the paradigm of deploying a robot-agent with general reasoning in a new home environment,where final fine-tuning and task-specific teaching are facilitated through natural language interactions between the user and the agent — akin to a homeowner explaining the nuances and details of a house to new tenants. Large Language Models-based task planners have significantly advanced robot task planning and human-robot interaction, enabling them to handle a wide range of real-world scenarios. However, language-conditioned policies are prone to failures due to the inherent complexity and constraints of real-world environments. In this work,we propose an agent-based framework that addresses plan failures by leveraging the capabilities of LLMs and real-time human feedback. The agent generates a goal scene graph and decomposes tasks into sub-goals to achieve the task objectives. When a failure occurs, the LLM-based framework revises the sub-goal sequence based on the goal scene graph, past experiences and/or human feedback to improve its performance for future tasks.

Keywords: Semantic Scene Understanding, Learning Categories and Concepts, Sensor Fusion
Abstract: The rise of competent mobile robots has introduced new requirements for environment mapping systems in recent years. Whereas 2D occupancy mapping was sufficient for most use cases in the past, capabilities like robust, legged locomotion paired with the availability of manipulators have increased the need for more task-specific and context-rich maps. More specifically, these maps need to contain additional semantic information to enable the efficient application of these robots. Typically, semantic maps are created based on ontologies specific to the planned application (e.g., modelling rooms, floors and target objects). We propose a more dynamic system that utilizes the inherent ontological knowledge within Large Language Models (LLMs) to build large, high-density semantic maps. Additionally, we also introduce a sensor fusion pipeline based on LLMs to account for the heterogeneous sensor setups and enable future usage with multi-robot teams. This work outlines the proposed pipeline and considerations regarding the design. The implementation and evaluation will be done as part of future publications.

Keywords: Humanoid and Bipedal Locomotion, Whole-Body Motion Planning and Control, Integrated Planning and Control
Abstract: Enhancing the robustness and speed of humanoid robot locomotion remains a key challenge in enabling these general-purpose robots to serve as effective tools for increasing industrial productivity. In this work, we propose a method for optimizing swing leg trajectories while considering hardware constraints and centroidal angular momentum to maximize walking speed. Additionally, to ensure contact stability, the center of mass trajectory is adapted to compensate for the induced angular momentum. We validated our approach using the humanoid robot TORO, setting a new walking speed record of 0.43 m/s for the system—an improvement of 16% over the previous record.

Keywords: Visual Tracking, Transfer Learning, Deep Learning for Visual Perception
Abstract: Tracking any point (TAP) in a video recently shifted the motion estimation paradigm from focusing on individual salient points with local templates to tracking arbitrary points with global image contexts. However, while research has mostly focused on driving the accuracy of models in nominal settings, addressing scenarios with difficult lighting conditions and high-speed motions remains out of reach due to the limitations of standard cameras. This work addresses this challenge with the first event camera-based TAP method. It leverages the high temporal resolution and high dynamic range of event cameras for high-speed tracking, and the global contexts in TAP methods to handle asynchronous and sparse event measurements. We further extend the TAP framework to handle event feature variations induced by motion, thereby addressing an open challenge in purely event-based tracking, with a novel feature alignment loss that ensures the learning of motion-robust features. Our method is trained using synthetic data from a new data generation pipeline. It shows strong cross-dataset generalization and performs 135% better on the average Jaccard metric than the baseline methods. Moreover, on an established feature tracking benchmark, it achieves a 19% improvement over the previous best event-only method and even surpasses the previous best events-and-frames method by 3.7%. The full paper is available at https://arxiv.org/pdf/2412.00133.

Keywords: Manipulation Planning, Mobile Manipulation, Optimization and Optimal Control
Abstract: Mobile manipulation of articulated objects demands advanced Task and Motion Planning (TAMP) strategies. Tasks such as opening doors or operating valves demand advanced planning strategies to manage complex contact phases and transitions. This challenge becomes even more complex in unstructured rescue scenarios. Articulated tracked robots with adjustable flippers are well-suited for rough terrain but introduce added complexity due to intricate robot–terrain interactions. Existing work focuses on quadrupedal robots or relies on the assumption of a flat ground. We propose a novel whole-body planner for mobile manipulation on uneven terrain. To our knowledge, it is the first approach specifically tailored to articulated tracked robots on non-flat ground. Our bi-level optimization employs an RRT-like search at the outer level to explore reference states and contact transitions, while an inner Optimal Control Problem (OCP) ensures feasibility. A penalty-based alternating direction method couples gradient-based manipulation optimization with a dynamic programming strategy for stability. We demonstrate that this planner can generate physically plausible trajectories for challenging door-opening and valve-turning tasks in unstructured scenarios.

Keywords: Human-Robot Teaming, Human-Robot Collaboration, Perception-Action Coupling
Abstract: The ability to anticipate human behavior is essential for robots to interact safely and efficiently with humans. In this work, we integrate anticipatory behavior into the control of a mobile manipulation robot using a smart edge sensor network. The external sensors provide global observations, future predictions, and goal information, enhancing the robot’s ability to navigate safely and collaborate effectively. We present two key approaches to human behavior anticipation: (1) safe navigation using projected human motion trajectories from the smart edge sensor network into the robot’s planning map, and (2) collaborative furniture handling, where the robot anticipates human intentions to achieve a predefined room layout. By incorporating human trajectories observed and predicted by the smart edge sensor network into the robot's planning framework, we enable it to benefit from global context information and thus navigate more safely in dynamic human-centered environments. In the collaborative furniture handling scenario, anticipation combines compliant control with goal inference, enabling efficient human-robot interaction.

Keywords: Haptics and Haptic Interfaces
Abstract: Grounded force-feedback (GFF) devices are a well-established and diverse category of haptic technology based on robotic arms. However, the number of designs and their specifications make it challenging to compare devices effectively. We address this challenge by presenting Haptify, a benchmarking system capable of evaluating GFF haptic devices in a thorough, fair, and non-invasive way. The user holds the instrumented device end-effector and moves it through a series of passive and active experiments. Haptify captures the interaction between the hand, device, and ground using a seven-camera optical motion-capture system, a custom 60-cm-square force plate, and a customized sensing end-effector. We propose six key metrics for evaluating GFF device performance: workspace shape, global free-space forces, global free-space vibrations, local dynamic forces and torques, frictionless surface rendering, and stiffness rendering. We then benchmark two commercial haptic devices using Haptify. The more expensive Touch X has a smaller workspace than the 3D Systems Touch, but it outputs smaller free-space forces and vibrations, smaller and more predictable dynamic forces and torques, and higher-quality renderings of a frictionless surface and high stiffness.

Keywords: Human-Robot Collaboration, Intention Recognition, Virtual Reality and Interfaces
Abstract: A major challenge in human-robot interaction is the mental model mismatch, which arises when a human's understanding of a robot's capabilities differs from the robot's actual operational model. Such mismatches can result in ineffective teaching, suboptimal performance, and interaction breakdowns. This project aims to quantify mental model mismatch by formalizing and comparing human expectations with robot learning processes, enabling a structured approach to improving teaching efficiency. By providing tailored feedback and enhancing transparency in the teaching process, the ultimate goal is to empower humans to form realistic expectations about robots and optimize instructional strategies. The vision is to foster intuitive and effective cooperative learning, where humans and robots collaborate seamlessly, leading to improved task execution and generalization capabilities across various scenarios.

Keywords: Modeling, Control, and Learning for Soft Robots, Model Learning for Control, Optimization and Optimal Control
Abstract: Soft robots are promising for various applications due to their intrinsic compliance and compact design. However, modeling and controlling such nonlinear systems is challenging, whereas learning-based approaches are ideally suited. For experimental validation, a real system is required, which should be open-source available to ensure comparability and reproducibility. In this paper, we present SPONGE as a possible benchmark platform. We give an overview of our recent research contributions where SPONGE was utilized for robot-based evaluations. Beyond soft robotics, the multi-input-multi-output (MIMO) system with up to 40 states and 20 inputs and several nonlinearities provides a test platform for real-world learning and control.

Keywords: Multifingered Hands, In-Hand Manipulation, Grasping
Abstract: The robustness of robotic hands is crucial for hand manipulation in real-world scenarios, as it ensures resilience to uncertainties and disturbances. Robotic hands with intrinsic compliance offer the advantage of optimized contact robustness and reduced reflected inertia, which significantly improves the cycle time for delicate objects. This paper presents Smarthand, a robust, compliant hand with a high force-to-weight ratio based on the DLR CLASH (three fingers) but with optimized kinematics for in-hand manipulation. The modular design allows for different hardware configurations, but also improves maintenance and calibration. The first tests on a two finger and three finger testbed showed new manipulation capabilities in compare to DLR Hand II, with significantly increased robustness and lower system costs.

Keywords: Neurorobotics, Sensor Fusion, Legged Robots
Abstract: Single-sensory robotic perception systems often suffer from high latency, redundant data processing, and limited viewpoints, reducing their responsiveness. This work proposes the fusion of event-based sensors—including vision, auditory, and tactile sensing—to develop a multimodal perception system. By integrating these modalities, quadruped robots can perceive their environment and make decisions within microseconds, enhancing their robustness and adaptability in complex terrains during autonomous tasks.

Keywords: Visual Learning, RGB-D Perception
Abstract: Autonomous robotic systems operating in dynamic indoor environments require robust perception capabilities to effectively navigate, interact, and reason about their surroundings. Central to achieving this is the development of general scene representations that capture both spatial and temporal dynamics, enabling real-time, robust, and long-horizon operations. This research aims to develop hybrid scene representations for scene deformation rendering, static-dynamic scene decomposition, and long-term dense correspondence. Inspired by the dual-process theory by Kahneman [1], we propose VIR4D, a multi-scale robot vision framework continually learns and adapts novel representations for the these tasks across different temporal scales, ultimately supporting robotic downstream tasks such as navigation, path planning, interaction and reasoning. By advancing the state-of-the-art in robot perception, this work seeks to enhance autonomy and robustness in complex human-centric environments, particularly in hospitals and workplaces, where safe and efficient robot operation is critical.

Keywords: Imitation Learning, Bimanual Manipulation, Visual Learning
Abstract: Imitation learning offers an efficient and intuitive way to teach humanoid robots new manipulation skills. However, acquiring generalizable task representations from sparse visual demonstrations remains challenging. This extended abstract summarizes our work on developing a novel Keypoints-based Visual Imitation Learning (KVIL) framework for humanoid robots. KVIL, a bottom-up approach, focuses on identifying invariant task features and extracting subsymbolic and symbolic task representations from scarce human demonstration videos. It comprises four key components: Uni-KVIL for unimanual task learning, Bi-KVIL for bimanual coordination, Seq-KVIL for learning action sequences, and Pro-KVIL for probabilistic task representation and inter-category generalization. These approaches collectively enable object-centric, viewpoint-invariant, and embodiment-independent task representations, allowing humanoid robots to generalize learned manipulation skills across object instances and categories. The framework is evaluated through real-world experiments, showcasing its effectiveness in learning daily tasks.

Keywords: AI-Enabled Robotics, Multi-Robot Systems, Intelligent Transportation Systems
Abstract: Transporting objects in household and warehouse environments is challenging due to varying object sizes, shapes, and weights. A common object transport strategy is pushing as it can be performed with relatively simple robots. To enable omnidirectional movement of the object through cluttered environments toward a goal position, we propose a novel framework for cooperative object pushing using a multi-robot system. This paper outlines our ongoing doctoral research project on a zero-shot end-to-end approach to object pushing in obstacle-cluttered environments. Our method does not require prior knowledge of the scene, obstacles, or the properties of the objects to be pushed. By leveraging the perception and reasoning abilities of vision language models, we identify obstacles and effective contact points for the robots to push the object along a planned path in a zero-shot manner. Our method is robust to random initial positions of the robots and can adapt to changes in the environment.

Keywords: SLAM, Object Detection, Segmentation and Categorization
Abstract: Autonomous robots must navigate dynamic environments where moving objects challenge traditional SLAM methods. Existing approaches struggle with efficiency, requiring extensive training data or high computational resources. We propose a lightweight dynamic SLAM system that detects and tracks moving objects directly in structured LiDAR point clouds, eliminating the need for pre-trained models or dense volumetric maps. By leveraging scan matching residuals, our method efficiently filters dynamic points. Evaluations on real-world and simulated datasets show significant gains in computational efficiency over state-of-the-art methods.

Keywords: Imitation Learning, Learning from Demonstration
Abstract: Designing modern imitation learning (IL) policies requires making numerous decisions, including the selection of feature encoding, architecture, policy representation, and more. As the field rapidly advances, the range of available options continues to grow, creating a vast and largely unexplored design space for IL policies. In this work, we present X-IL, an accessible open-source framework designed to systematically explore this design space. The framework's modular design enables seamless swapping of policy components, such as backbones (e.g., Transformer, Mamba, xLSTM) and policy optimization techniques (e.g., Score-matching, Flow-matching). This flexibility facilitates comprehensive experimentation and has led to the discovery of novel policy configurations that outperform existing methods on recent robot learning benchmarks. Our experiments demonstrate not only significant performance gains but also provide valuable insights into the strengths and weaknesses of various design choices. This study serves as both a practical reference for practitioners and a foundation for guiding future research in imitation learning.

Keywords: Imitation Learning, Learning from Demonstration
Abstract: Learning for manipulation requires using policies that have access to rich sensory information such as point clouds or RGB images. Point clouds efficiently capture geometric structures, making them essential for manipulation tasks in imitation learning. In contrast, RGB images provide rich texture and semantic information that can be crucial for certain tasks. Existing approaches for fusing both modalities assign 2D image features to point clouds. However, such approaches often lose global contextual information from the original images. In this work, we propose a novel imitation learning method that effectively combines the strengths of both point cloud and RGB modalities. Our method conditions the point-cloud encoder on global and local image tokens using adaptive layer norm conditioning, leveraging the beneficial properties of both modalities.

Keywords: Multi-Robot Systems, Human Detection and Tracking, Optimization and Optimal Control
Abstract: Securing large industrial facilities against intruders demands more sophisticated solutions than traditional static sensors or manual patrols can provide. This paper outlines a novel framework for coordinating heterogeneous robot teams consisting of aerial and ground robots for active target search and tracking in industrial surveillance applications. The proposed approach combines probabilistic state estimation through particle filtering with human motion prediction using the Social Force Model. Unlike existing methods that treat targets as simplified entities, this framework explicitly models human motion patterns and environmental constraints. The system leverages complementary strengths of aerial and ground robots while incorporating navigation graphs for environment-specific constraints.

Keywords: AI-Enabled Robotics, Failure Detection and Recovery, Agent-Based Systems
Abstract: Autonomous mobile robots (AMRs) and mobile manipulators are critical components in modern logistics and manufacturing. However, their efficiency depends on the timely detection and resolution of anomalies to prevent downtime. Traditional predictive maintenance systems often require large data sets and cloud connectivity, which makes them unsuitable for some environments. The IntelliVoiceAnalytic project addresses these challenges by developing an onboard AI-based anomaly detection system that can operate with minimal training data. The solution integrates offline-capable speech recognition and synthesis tools to enable intuitive speech-based human-robot interaction. This paper presents the project approach, including data acquisition, synthetic anomaly generation, and development of a demonstrator AMR.

Keywords: RGB-D Perception, SLAM
Abstract: Efficient 6D object pose estimation is essential for applications such as augmented reality, autonomous driving, and robotics. While model-based methods show promise, model-free approaches incur high computational costs for rendering and pose inference in live RGB-D video. To address this, we introduce 6DOPE-GS, an online 6D pose estimation and tracking method that utilizes a single RGB-D camera and Gaussian Splatting. By leveraging fast differentiable rendering, 6DOPE-GS optimizes both 6D poses and 3D reconstruction. Our approach employs incremental 2D Gaussian Splatting with dynamic keyframe selection for robust tracking and an opacity-based pruning mechanism for stable and efficient training. Evaluations on the HO3D and YCBInEOAT datasets show that 6DOPE-GS matches state-of-the-art performance while achieving a 5x speedup, demonstrating its suitability for real-time dynamic object tracking and reconstruction.

Keywords: Task Planning, Manipulation Planning
Abstract: The MoveIt Task Construction (MTC) framework provides a modular and fine-grained approach for practical multi-modal planning. Tasks define and plan high-level manipulation actions with interdependent components, such as pick&place or handover tasks. Hierarchical organization of basic planning stages is supported through container structures, allowing for sequential as well as parallel compositions. The flexibility of the framework is illustrated in multiple scenarios performed on various robot platforms, including bimanual ones.

Keywords: SLAM, Data Sets for SLAM
Abstract: CHAD TSDF is a novel representation for Trun- cated Signed Distance Fields (TSDFs) that maintains the associ- ation between scan poses and their corresponding TSDF values. This association is crucial for applications in Simultaneous Localization and Mapping (SLAM) to enable pose graph optimization post map update. We evaluate CHAD TSDF on the MulRan sequences and a large-scale dataset created with a mobile scanning device and compare it with VDBFusion in memory efficiency and insertion performance.

Keywords: Machine Learning for Robot Control, Manipulation Planning, Reinforcement Learning
Abstract: Pushing objects to target positions is an important skill for robots. However, the task is challenging due to unknown physical properties. We improve a state-of-the-art model by introducing a new memory-based vision-proprioception reinforcement learning (RL) model to push objects more precisely using fewer corrective movements.

Keywords: Human-Robot Collaboration, Cooperating Robots, Emotional Robotics
Abstract: This abstract gives an overview of the structure and the achievements of the European MindBot project in applying a combination of socially interactive agents (SIAs) and collaborative robots (cobots) in an industrial working scenario.

Keywords: Data Sets for Robot Learning, AI-Enabled Robotics, Imitation Learning
Abstract: We proposed a framework to learn the object placement strategy in transformer-based simulation environments. The model is trained to learn the optimal object rearrangement and placement by encoding object properties, spatial relationships, and task dependencies into embedding. We evaluate our approach in the Multiverse simulator environment and use the objects as their primitives. The transformer model demonstrates adaptability to unseen scenarios, adherence to sequential constraints, and robustness in dynamic environments. This approach has a wide range of applications in robotic manipulation tasks, such as automated storage and retrieval systems and house holds environment.

Keywords: Learning from Demonstration, Semantic Scene Understanding, Force and Tactile Sensing
Abstract: The industrial sector continuously demands efficient methods to enable non-expert operators to reprogram robots in a timely and cost-effective manner. Advances in task-level programming (TLP), robotic skill acquisition, and Learning from Demonstration (LfD) have yielded promising outcomes. Nonetheless, many existing approaches remain dependent on extensive datasets or necessitate prior user expertise in robotic systems. This paper introduces a framework for deriving parameterized skill sequences from passive observation of human demonstrations. These skill sequences reflect human behavior and enable the design of a task plan to execute on the robot. Since passive observation alone does not provide information about the physical properties of objects, which are critical for effective manipulation, our approach integrates robotic tactile and kinesthetic sensing to estimate both static and dynamic physical properties of the manipulated objects.

Keywords: Software Tools for Robot Programming, AI-Based Methods, Force Control
Abstract: Intuitive robot programming approaches, such as Programming by Demonstration, enable non-experts to teach robots tasks through hand-guided demonstrations. However, for in-contact tasks requiring force control for parts of the motion, automated segmentation of motion types is crucial for improving usability. Considering the related work, this work proposes a neural network-based segmentation method categorizing motion into three fundamental types using input data of cartesian twists and wrenches at the end effector. We collect demonstration data using a 7-axis robot with a wrist-mounted force torque sensor and generate ground truth labeling by hand. While still in development, our research direction aims to enable non-experts to program in-contact tasks more intuitively.

Keywords: Cognitive Control Architectures, AI-Enabled Robotics, Software Tools for Robot Programming
Abstract: CLIPS is a rule-based programming language for building knowledge-diven applications. It is well suited for the complex task of coordinating autonomous robots, hence would be a great addition to ROS framework for open-source robotics applications. Inspired by the CLIPS Executive originally developed for the lesser known Fawkes robotics framework, we present an Integration of CLIPS into the ROS ecosystem.

Keywords: SLAM, Mapping, Localization
Abstract: Simultaneous Localization and Mapping (SLAM) enables mobile robots to navigate autonomously without relying on external positioning systems. In scenarios with obstructed vision caused by rain or snow, radar has emerged as a reliable sensing modality due to its resilience to environmental particles. Novel 4D imaging radars provide 3D geometric and velocity data and have shown remarkable performance in odometry estimation. However, the sparsity of measurements and limited field of view (FOV) presents challenges for detecting loop closures, particularly from reverse viewpoints with minimal scan overlap. This work investigates using 4D radar for loop closure in SLAM. Specifically, submaps are employed to alleviate sparsity – and together with introspective quality measures – mitigate false detections in feature-degenerate areas. Experimental results highlight the approach’s effectiveness in improving trajectory estimation and achieving robust loop closure detection across diverse geometric settings.

Keywords: Sensor Networks, Environment Monitoring and Management, Software, Middleware and Programming Environments
Abstract: Monitoring of gas emissions is critical for ensuring safety, efficiency, and compliance in industrial plants. Approaches based on stationary sensor networks leave areas of the plant unmonitored and lack adaptability to changing conditions. Combining stationary with mobile sensors on quadruped robots offers a robust approach with improved spatial resolution, adaptability and responsiveness. To streamline the integration of such systems, a unified and open software framework is required. We propose to abstract the low-level implementations, e.g., the robot control layers, using the REST API to facilitate the implementation of high-level tasks such as data visualization and AI-based analysis, thus enabling scalable monitoring solutions. An indoor gas release experiment was conducted, which showed the benefits of the quadruped robot-based and the combined sensing approach.

Keywords: Marine Robotics, Field Robots, Motion Control
Abstract: Over the past two decades, autonomous under-ice exploration has been increasingly catching attention, particularly in the quest to discover life beneath the surfaces of icy moons. One of the main directions research has been going is on designing an exploration methodology using Autonomous Underwater Vehicles (AUVs). As part of an effort to address these challenges, the "Europa-Explorer" project line concentrated on conducting a pilot survey for a future mission to explore the Jovian moon. The project aimed to validate the proposed exploration scenario through analogue simulations in under-ice environments on Earth. Throughout this project line, researchers and engineers at the German Research Center of Artificial Intelligence (DFKI) have developed an AUV "DeepLeng", as a system dedicated to sub-ice navigation to test the proposed mission. This work showcases the experiences made throughout the past 3 years with the system, providing (1) a brief description of the AUV, (2) its localization and docking algorithms, (3) system controls and trajectory optimization, and (4) the field trials conducted in an ice-covered in Abisko, Sweden.

Keywords: Human-Robot Collaboration, Virtual Reality and Interfaces, Physical Human-Robot Interaction
Abstract: This paper presents a Virtual Reality-based control tool for Wizard of Oz experiments, designed to bridge the gap between VR simulations and real-world robotics applications. The study evaluates the tool's effectiveness in both environments, focusing on feature transition, scalability, and real-world constraints such as sensor feedback and latency. The results highlight the tool's potential to support both VR-based and real-world robot testing, while also identifying challenges such as control dynamics and localisation. Future work will explore improvements in motion realism, sensor integration, and AI-driven interaction models to increase the tool's scalability and applicability to different robot types.

Keywords: Art and Entertainment Robotics, Swarm Robotics, AI-Enabled Robotics
Abstract: The use of foundation models, such as large language models (LLMs), in robotics has introduced new possibilities for interactions and programming, but their integration into safety-critical systems remains challenging. In this extended abstract we outline how model predictive control (MPC)-based filters from control theory can help bridge this gap to allow the deployment of foundation models in complex robotic systems. Specifically, we let an LLM design the motions of a swarm for drone performances and directly deploy the filtered output without any manual intervention on real robots. In our experiments, we show how choosing suitable action spaces such as motion primitives for the LLM enables us to scale towards large swarms of up to 20 drones.

Keywords: Autonomous Vehicle Navigation, Planning under Uncertainty, Machine Learning for Robot Control
Abstract: This study presents a navigation algorithm tailored for ground vehicles in challenging 3D off-road terrains. Off-road navigation shows great potential in diverse applications, yet many challenges remain, mainly attributed to uncertainties in vehicle-terrain interactions. The proposed algorithm addresses such intricate interactions by exploiting 3D vehicle dynamics and exploring different terrain conditions, all while accounting for inherent uncertainties. First, learning from driving data, our algorithm extracts traversability features via neural networks, and these features are fed into the Gaussian process regression (GPR) to model uncertain vehicle-terrain interactions and quantify the corresponding uncertainties. The quantified uncertainties are seamlessly integrated into the evaluation of traversability costs, yielding a path-planning strategy that attunes both the safety and agility of vehicle maneuvering. Our approach has been rigorously tested in off-road simulation environments, demonstrating its real-time feasibility and robustness.

Keywords: Rehabilitation Robotics, Physical Human-Robot Interaction, Human-Centered Robotics
Abstract: Training with adaptive task difficulty can account for individual needs, provide challenging tasks, and adapt to the user's performance. Automating this labor-intensive process is especially important in the face of a rapidly aging population and a shortage of healthcare professionals. The main challenges are automatic assessment of user performance, adaptive task difficulty to maintain intensity and motivation, and dynamic exercise scheduling. The goal of this extended abstract is to highlight the research gap in this area and to present a framework that could help to realize individualized lower-limb motor training with the robotic walker RoboTrainer. The proposed framework RoboTrainerAID uses machine learning techniques for automatic assessment and spatial control actions for adaptive task difficulty. A comprehensive evaluation will be performed in a future user study.

Keywords: Environment Monitoring and Management, AI-Based Methods, Semantic Scene Understanding
Abstract: In recent years, mobile walking robots such as ANYmal and Spot, combined with powerful inspection sensor payloads, have emerged as a promising solution to reduce the impact of labor shortages in monitoring and inspection tasks. Current robot inspection systems are still constrained to pre-defined missions, conducting measurements at fixed points in the monitored environment and requiring in-domain knowledge to ensure feasible mission design. By expanding the system's capabilities, we want to enable continuous intelligent monitoring missions where the system can identify areas of high potential information gain in the monitored space and implicitly set goals and tasks. One crucial step towards higher levels of autonomy is the detection and localization of anomalies. This paper proposes a pipeline deployed on an off-the-shelf ANYmal D robot to detect, localize, and map sound anomalies as a crucial domain for predictive maintenance and general monitoring tasks. To detect anomalies, we focus on adapting methods used in the field of Anomalous Sound Detection (ASD) to a mobile robot use case. For sound localization, we leverage the robot's mobility to triangulate sound source positions. Combining the data with the positional information, we propose a method to save the data into a 3D voxel-based map for spatio-temporal reconstruction further down the line.

Keywords: Task and Motion Planning, Whole-Body Motion Planning and Control, Optimization and Optimal Control
Abstract: This work presents preliminary results on an optimization based task and motion planning (TAMP) method for loco-manipulation systems. Our method prioritizes maintaining long term dependencies in the optimization in order to enable planning for complex behaviors. In our bilevel formulation the outer discrete layer employs a sampling-based graph search to explore symbolic transitions, while the inner continuous layer refines these via whole body dynamics constraints as well as kinematic and dynamic contact interaction constraints. Preliminary results on a quadruped loco-manipulation task demonstrate that our method naturally produces complex behaviors, allowing us to exploit the dynamic capabilities of our robots in loco-manipulation settings.

Keywords: Vision-Based Navigation, Autonomous Vehicle Navigation, AI-Enabled Robotics
Abstract: Effective robot navigation in dynamic environments is a challenging task that depends on generating precise control actions at high frequencies. Recent advancements have framed navigation as a goal-conditioned control problem. Current state-of-the-art methods for goal-based navigation, such as diffusion policies, either generate sub-goal images or robot control actions to guide robots. However, despite their high accuracy, these methods incur substantial computational costs, which limits their practicality for real-time applications. Recently, CFM framework has emerged as a more efficient and robust generalization of diffusion. In this work, we explore using CFM to learn action policies that help the robot navigate its environment. Our results demonstrate that CFM can generate highly accurate robot actions. CFM not only matches the accuracy of diffusion policies but also significantly improves runtime performance. This makes it particularly advantageous for real-time robot navigation, where swift, reliable action generation is vital for collision avoidance and smooth operation.

Keywords: Constrained Motion Planning, Intelligent and Flexible Manufacturing, Optimization and Optimal Control
Abstract: Industrial Robots are vital for modern automated manufacturing, but they struggle in applications with strong requirements for absolute accuracy or stiffness. Good task performance is only achievable by optimizing any possible functional redundancy. This abstract proposes a sampling-based approach for optimizing robot trajectories over long tool paths in complex environments. Limited initial evaluation is performed to discuss general feasibility.

Keywords: Human-Centered Robotics, Social HRI, Robot Safety
Abstract: In current research and service industries social robots are gaining relevance; however, their systematic integration across all phases of the Customer Journey (CJ) is predominantly focused to initial stages. The inclusion of social robots in all phases presents opportunities for enhancing customer experience (CX), e.g., through spatial integration and adapting types of service offerings. However, there is a lack of methodologies for their strategic deployment in CX. To address this, the authors developed the Human-Robo Journey (HRJ) framework, which adapts the CJ to focus on consumers’ behavior with social robots in service industries. This paper introduces a holistic multi-level approach integrating Customers (1:Persona Profile Development), Service Setting (2:Experience Environment), and CJ (3:Human-Robo Journey). Developed through a systematic literature review, and interdisciplinary workshops on CX and CJ Mapping (n=30), the framework was tested in a field study in German libraries (n=65). Libraries offer a conducive environment for diverse audiences, providing transferable insights applicable across service industries. The paper contributes to technology-driven behavioral insights, highlighting the potential of social robots to enhance service delivery throughout the CJ.

Keywords: Deep Learning for Visual Perception, Perception-Action Coupling
Abstract: Robots need to perform tasks by interacting with and manipulating objects in their environment. Affordance prediction involves predicting the correct regions of objects to interact with to perform a task. Existing methods simplify this problem to naive object part segmentation. We propose a framework to extract and predict `real' and `actionable' affordances from videos of humans performing or demonstrating tasks. We propose the 2HANDS dataset of actionable affordances & show the performance of a baseline model to learn meaningful and actionable affordances.

Keywords: Integrated Planning and Learning, Manipulation Planning, Data Sets for Robot Learning
Abstract: Generating diverse samples under hard constraints is a core challenge in many areas. In this work we discuss NLP Sampling as an integrative view and framework to combine methods from the fields of MCMC, constrained optimization, as well as robotics. While NLP Sampling is a general method for constrained sampling, we demonstrate it here for sampling manipulation strategies. We discuss such model-based diverse manipulation sampling as an alternative to data collection via human demonstration.

Keywords: Marine Robotics, Mechanism Design, Field Robots
Abstract: We introduce the SPiraling Intelligent Robotics Underwater monitoring pLAtform (SPIRULA) as a a novel un- derwater monitoring platform designed to reduce the logistical effort, cost, and complexity of long-term seafloor observations. By combining a static lander with a mobile autonomous vehicle tethered to it, the SPIRULA design enables sharing of data and energy, resulting in a robust, compact and simplified system. By unwinding and winding from a passive drum with a taught tether the SPIRULA vehicle is forced on a circle involute path around the SPIRULA lander, allowing for repeatable and efficient coverage of a disc around the lander. With a 20 m tether monitoring of an area of 1200 m2 is possible. Initial trials in a research pool demonstrate the platform’s potential for long-term monitoring in marine biology, chemistry, and geology.

Keywords: Virtual Reality and Interfaces, Telerobotics and Teleoperation, Learning from Demonstration
Abstract: This paper presents IRIS, an Immersive Robot Interaction System designed for robot data collection and interaction through Extended Reality (XR). Unlike existing XR-based systems, which are difficult to reproduce and tied to specific simulators, IRIS offers a flexible, extensible framework supporting multiple simulators, benchmarks, and headsets. An unified scene specification is generated directly from simulators or real-world sensors and transmitted to XR headsets, creating identical scenes in XR. This specification allows IRIS to support any of the objects and robots from simulators. With features like shared spatial anchors and robust communication protocols, IRIS facilitates collaborative data collection across multiple users. IRIS has been tested on Meta Quest 3 and HoloLens 2, and it showcased its versatility across a wide range of real-world and robot simulators such as MuJoCo, IsaacSim, CoppeliaSim, and Genesis. A user study on the LIBERO benchmark shows that IRIS significantly outperforms the baseline in both objective and subjective metrics, highlighting its potential to advance data collection and interaction in robotics and robot learning.

Keywords: Software Architecture for Robotic and Automation, Software-Hardware Integration for Robot Systems, Multi-Robot Systems
Abstract: We provide algorithms and software to track teams of robots with a motion-capture system. Unlike the vendor-specific software, our approach can use identical marker configurations on each robot, and supports cases where only the position, not the full pose, is required. Our approach is available as open-source standalone and as a ROS 2 package.

Keywords: Force and Tactile Sensing, RGB-D Perception, Sensor Fusion
Abstract: This paper presents an approach to robotic in-hand object pose estimation, combining visual and tactile information to accurately determine the position and orientation of objects grasped by a robotic hand. We address the challenge of visual occlusion by fusing visual information from a wrist-mounted RGB-D camera with tactile information from vision-based tactile sensors mounted on the fingertips of a robotic gripper. Our approach employs a weighting and sensor fusion module to combine point clouds from heterogeneous sensor types and control each modality's contribution to the pose estimation process. We use an augmented Iterative Closest Point (ICP) algorithm adapted for weighted point clouds to estimate the 6D object pose. Our experiments show that incorporating tactile information significantly improves pose estimation accuracy, particularly when occlusion is high. Our method achieves an average pose estimation error of 7.5 mm and 16.7 degrees, outperforming vision-only baselines by up to 20%. To validate the practical applicability of our method, we conducted an insertion task experiment, demonstrating the ability to perform precise object manipulation in a real-world scenario.

Keywords: Autonomous Agents, Deep Learning for Visual Perception, Data Sets for Robot Learning
Abstract: Autonomous agents need to precisely localize themselves in diverse environments, including in GPS-denied indoor environments. One approach for precise localization is visual place recognition (VPR), which estimates the place of an image based on previously seen places. State-of-the-art VPR models require high amounts of memory, making them unwieldy for mobile deployment, while more compact models lack robustness and generalization capabilities. This work overcomes these limitations for robotics using a combination of event-based vision sensors and a guided variational autoencoder (VAE). The spiking neural network encoder part of our model is compatible with power-efficient, low latency neuromorphic hardware. The VAE successfully disentangles the visual features of 16 places in our new indoor VPR dataset with a classification performance comparable to other state-of-the-art approaches. When tested with visual inputs from unknown scenes, our model can distinguish between these places, which demonstrates a high generalization capability. Our compact and robust guided VAE poses a promising model for visual place recognition that can significantly enhance mobile robot navigation in known and unknown indoor environments.

Keywords: Mapping, In-Hand Manipulation, Deep Learning Methods
Abstract: We propose a dynamic reconstruction algorithm for consistent canonical space representation, conditioned on hand poses or temporal embeddings as input. Our method learns an interaction model, representing a digital interaction object, from RGB-D Hand-Object-Interaction (HOI) videos captured from either an egocentric or third-person perspective.

We employ a network-based dynamic Gaussian Splatting reconstruction framework, decomposing Gaussian transformations into global and local motions. Sequence-level supervision enables the training of both networks separately or jointly, depending on the sequence type. To ensure consistent alignment over consecutive frames, we supervise the reconstruction with long-term keypoint tracks. We also address hand-object occlusions by controlling Gaussian supervision in occluded hand regions. We showcase our current progress, qualitative comparisons, and future directions, highlighting potential robotics applications of our approach.

Keywords: SLAM
Abstract: Visual Odometry (VO) and SLAM are fundamental components for spatial perception in mobile robots. Despite enormous progress in the field, current VO/SLAM systems are limited by their sensors’ capability. Event cameras are novel visual sensors that offer advantages to overcome the limitations of standard cameras, enabling robots to expand their operating range to challenging scenarios, such as high-speed motion and high dynamic range illumination. We propose a novel event-based stereo VO system by combining two ideas: a correspondence-free mapping module that estimates depth by maximizing ray density fusion and a tracking module that estimates camera poses by maximizing edge-map alignment. We evaluate the system comprehensively on multiple real-world datasets, spanning a variety of camera types (manufacturers and spatial resolutions) and scenarios (driving, hand-held, egocentric, etc). The quantitative and qualitative results demonstrate that our method outperforms the state of the art in most of the test sequences by a margin. Project page: https://github.com/tub-rip/ES-PTAM

Keywords: Telerobotics and Teleoperation, AI-Enabled Robotics, Bimanual Manipulation
Abstract: Manipulating deformable linear objects (DLOs) such as ropes is challenging due to their complex dynamics. To address these issues, we present a novel assistive teleoperation framework that combines human expertise with autonomous assistance. Our approach integrates a vision-based module to identify grasp poses, a shared autonomy mechanism that balances human input with autonomous guidance, and an optimization-based inverse kinematic solver for smooth, collision-free manipulation. Additionally, a virtual reality (VR) interface provides intuitive control and real-time feedback to the operator. A user study on knot untangling under time-delayed and non-delayed conditions shows that shared autonomy enhances task performance under delay while reducing the operator's physical and mental workload. These findings highlight the potential of shared autonomy to improve teleoperation systems for complex DLO manipulation, particularly in environments affected by communication delays or uncertainties.

Keywords: Human-Robot Collaboration, Virtual Reality and Interfaces, Simulation and Animation
Abstract: This extended abstract presents a novel research direction: the evaluation of human-robot collaboration for industrial assembly processes via mixed reality methods. Integrating collaborative robots into manufacturing processes has transformed industry dynamics, enhancing efficiency and adaptability. However, traditional training methods for human-robot collaboration have struggled to keep pace with rapid technological advancements. Our research proposal explores the potential of Mixed Reality to revolutionize employee training, offering immersive and interactive environments that blend digital and physical realities. We aim to optimize human-cobot processes by creating a Mixed Reality environment that simulates collaborative tasks. This project assesses its efficacy as a training tool compared to traditional methods, focusing on learning outcomes and engagement.

Keywords: Telerobotics and Teleoperation, Surgical Robotics: Planning, Networked Robots
Abstract: This paper introduces a novel 5G-enabled telesurgical platform designed to enhance remote surgery through advanced robotic control and tactile feedback. The system integrates three Franka serial arms equipped with custom modular drive unit, providing precise and adaptive motion of an attached surgical tool. The platform leverages highspeed communication networks, including 5G, to facilitate realtime haptic feedback and AI-assisted control strategies for remote surgical applications. By advancing robotic dexterity, force feedback, and open control framework, this platform aims to advance robotic and communication research aspects in telesurgery. The performance of the system and different features are shown in accompanying video.

Keywords: Imitation Learning, Data Sets for Robot Learning
Abstract: A central challenge in developing robots that can relate human language to their perception and actions is the scarcity of natural language annotations in diverse robot datasets.

We introduce NILS: Natural language Instruction Labeling for Scalability. NILS automatically labels uncurated, long-horizon robot data at scale in a zero-shot manner without any human intervention by combining pretrained vision-language foundation models.

Evaluations on BridgeV2, Fractal, and a kitchen play dataset show that NILS can autonomously annotate diverse robot demonstrations of unlabeled and unstructured datasets. We use NILS to label over 115,000 trajectories obtained from over 430 hours of robot data.

We open-source our auto-labeling code and generated annotations on our website: http://robottasklabeling.github.io.

Keywords: Legged Robots, Localization, Sensor Fusion
Abstract: Robust and accurate proprioceptive state estimation of the main body is crucial for legged robots to execute tasks in extreme environments where exteroceptive sensors, such as LiDARs and cameras may become unreliable. In this paper, we propose a state estimation system for the legged robots main body that fuses the measurements from the body-mounted IMU (Body-IMU), joint encoders, and multiple leg-mounted IMUs (Leg-IMU) using an extended Kalman filter (EKF). The filtering system contains the states of all IMU frames. Moreover, the Leg-IMUs are used to detect foot contact, thereby providing zero velocity measurements to update the state. Additionally, we incorporate the relative position constraints between the Body-IMU and Leg-IMUs to improve the pose estimation accuracy.

Keywords: Telerobotics and Teleoperation, Distributed Robot Systems, Compliance and Impedance Control
Abstract: The design of an IEEE Standard for Haptic Codecs for the Tactile Internet IEEE 1918.1.1-2024 [1] was initiated in 2016 to analyze the best-performing technologies in haptics, data transmission and teleoperation. An international consortium collected and evaluated the most promising approaches. Thereby, the teleoperation methods were compared based on metrics as position tracking quality and force transparency. In 2024, as part of IEEE 1918.1.1-2024, the first IEEE standard for delayed teleoperation which is the focus of this document was published by the IEEE standards association. The core methodologies of this standard are the Deadband-Based Haptic Packet Rate Reduction Approach (DB approach [2]) and the Energy Reflection-Based Time-Domain Passivity Approach (TDPA-ER [3]). The respective combination was first published in [4]. This document summarizes recent extensions of the IEEE Standard further improving the quality of control forces, enhancing safety and the physical realism of the coupling. Thanks to the design of the teleoperation control based on the network representation and passivity, the different methods are highly modular allowing for simplified extensions of the IEEE Standard framework. Thus, the DB approach can be simply considered as part of the communication channel in Fig. 3 presenting the network representation of the standard. The figure depicts the coupling controller with stiffness and damping element and the energy monitoring unit (MU) an passivity controllers PC as the core parts of TDPA-ER. Elements additional to the standard are presented in red color. The bold arrows indicate the energy flow in the system which is the main reference for the passivity control in TDPA-ER. The MU collects the energy input from input device (Port 1) and robot (Port 6) and distributes this available energy to the outputs at the same ports. Thereby, the passivitiy controllers ensure via attenuation of the control forces that only available energy exits the network and that the passivity criterion is not violated. The modularity visible from Fig. 3 also enabled the extension of the TDPA-ER to 6-DoF [5] and to the position-position architecture [6].

Keywords: Human-Aware Motion Planning, Human Detection and Tracking, Datasets for Human Motion
Abstract: In this paper we present a hierarchical motion and intent prediction system prototype, designed to efficiently operate in complex environments while safely handling risks arising from diverse and uncertain human motion and activities. Our system uses an array of advanced cues to describe human motion and activities, including generalized motion patterns, full-body poses, heterogeneous agent types and causal contextual factors that influence human behavior.

Keywords: Intelligent and Flexible Manufacturing, Industrial Robots, Motion and Path Planning
Abstract: Skill-based production enhances agility and efficiency in modern manufacturing by enabling modularity, interoperability, and dynamic reconfiguration, thereby reducing setup times and operational complexity. This paper presents our developed skill-based applications in industrial robotics for tasks such as sorting, disassembly, machining, and intralogistics.

Keywords: AI-Enabled Robotics
Abstract: The Virtual Research Building (VRB) is a next-generation platform, representing a transformative approach to advancing research, education, and industry collaboration within the robotics and artificial intelligence domains. Leveraging state-of-the-art digital infrastructure such as semantic digital twins (semDTs), containerized software ecosystems, and scalable cloud-based infrastructure, the VRB provides a comprehensive virtual environment that integrates simulations, remote accessibility, and support for research and development. This environment enables researchers and industry partners to conduct experiments, prototype designs, test cutting-edge technologies and deploy robotics systems in a controlled and cost-effective manner, fostering innovation, accessibility, and global collaboration. By reducing hardware dependencies and enabling reproducible experiments, the VRB significantly accelerates research workflows, supports modular robot integration, and facilitates rapid validation of algorithms and software.

Keywords: Force and Tactile Sensing, Deep Learning in Grasping and Manipulation, Sensor-based Control
Abstract: Optical tactile sensors provide high spatial resolution, but struggle to offer fine temporal resolutions. To overcome this shortcoming, we study the idea of replacing the RGB camera with an event-based camera and introduce a new event-based optical tactile sensor called Evetac. Along with hardware design, we develop touch processing algorithms to process its measurements online at 1000 Hz. We devise an efficient algorithm to track the elastomer’s deformation through the imprinted markers despite the sensor’s sparse output. Benchmarking experiments demonstrate Evetac’s capabilities of significantly reducing data rates compared to RGB optical tactile sensors. Moreover, Evetac’s output and the marker tracking provide meaningful features for learning data-driven slip detection models. The learned models form the basis for a robust and adaptive closed-loop grasp controller capable of handling a wide range of objects. We believe that fast and efficient event-based tactile sensors like Evetac will be essential for bringing human-like manipulation capabilities to robotics.

Keywords: Social HRI, Human-Centered Robotics, Human-Robot Teaming
Abstract: Mental healthcare is a promising field of application for Socially Assistive Robots (SARs). However, in this context, most research has been targeted at the treatment of autism spectrum disorder and dementia. In this extended abstract, we give an overview of the SPARCi project, with which we aim to investigate the efficacy of SARs in the treatment for internalizing disorders.

Keywords: Sensor Fusion, Recognition, Gesture, Posture and Facial Expressions
Abstract: Hand gesture recognition through force myography (FMG) is a valuable approach for human-computer interaction and assistive technologies. This paper presents a hybrid hand gesture recognition system that integrates forearm-based FMG with direct finger motion sensing. The system uses an eight-channel FMG band to capture muscle deformations and a sensorized glove with four strain sensors to monitor individual finger movements. To optimize feature selection and reduce computational complexity, Recursive Feature Elimination (RFE) and Feature Importance (FI) methods were applied, revealing that glove strain sensors, particularly those placed on the index and pinky fingers, are most informative for gesture classification. Experimental results show that selective channel optimization significantly improves recognition performance, offering promising applications in prosthetic control, rehabilitation, and advanced human-robot interaction.

Keywords: Imitation Learning, Deep Learning in Grasping and Manipulation, Visual Learning
Abstract: This work introduces FLOWER, an efficient, open-source Vision-Language-Action policy that outperforms current generalist and specialist policies while significantly reducing computational demands for pretraining and inference. By combining a Rectified Flow Policy for expressive multimodal action generation with a compact Vision-Language Model for robust semantic grounding, FLOWER achieves high performance with less reduced training time and memory. Evaluations on 5 simulated benchmarks and over 190 real-world tasks demonstrate its superiority over foundation policies like OpenVLA across diverse control tasks. The full pretraining and fine-tuning code, along with trained weights, will be open-sourced to promote further research and accessible deployment.

Keywords: Grasping, Human Factors and Human-in-the-Loop, Learning from Demonstration
Abstract: Perception-based robotic grasping is a fundamental ability for service robots to interact with objects in unstructured environments like households. For many manipulation tasks, objects need to be grasped suitably considering the prospective manipulation task. As opposed to the generic grasp pose estimation, task-oriented grasp pose estimation requires cognitive reasoning and task knowledge to yield a correct grasp configuration, in addition to estimating a stable grasp pose. Exploiting top-down human knowledge during task-oriented grasp planning can effectively mitigate these challenges. With this aim, we design a human-in-the-loop grasping system based on remote grasp area demonstrations. We create an interactive segmentation interface that facilitates demonstrations by non-experts. Following the online support from the human for the task-oriented grasp area, the system plans grasp poses in the demonstrated region and the robot grasps the object autonomously. After multiple demonstration sessions for a specific task, we deploy a few-shot learning method and an image segmentation network to estimate the task-oriented grasp area on a new scene without a human intervention. Experimental results indicate a successful learning of the task-oriented grasp areas after 5-shot training. This enables the system to learn new task-oriented grasping skills by human support and to improve its autonomy in time.

Keywords: Rehabilitation Robotics, Prosthetics and Exoskeletons, Transfer Learning
Abstract: For more effective sensorimotor post-stroke rehabilitation, traditional physiotherapy should be combined with robot-assisted therapy. To this end, the RECUPERA exoskeleton was developed to enhance the available therapy options and to reduce the burden on therapists by supporting upper-body movements during therapy sessions. To enable an intuitive and natural interaction with the exoskeleton, a virtual reality-based rehabilitation scenario was designed, in which the user’s movement intentions were decoded through continuous processing of EEG and EMG signals. In this approach, two neural network models were trained for robust online detection of reaching movement intentions from the EEG. The outputs of these models were then combined into an ensemble model and a classifier transfer approach was applied additionally. This transfer approach utilized the mirror mode functionality of the RECUPERA exoskeleton to mirror the movements of the unaffected arm onto the affected arm and drastically reduced calibration times. Additionally, grasping actions were decoded with the help of EMG signals. Finally, we successfully demonstrated the effectiveness of our approach in practice on a healthy subject by integrating online decoding of EEG and EMG signals with the mirror-mode functionality of the exoskeleton in an interactive virtual reality-based rehabilitation scenario.

Keywords: Humanoid Robot Systems, Reinforcement Learning, Simulation and Animation
Abstract: Humanoid robotics has made significant progress in vision, manipulation, locomotion, yet upper-body capabilities remain underexplored. This study addresses this gap by focusing on the challenging chin-up task, which requires a high strength-to-mass ratio in humans.

Our experiments were conducted using the RH5 humanoid, a whole-body system with a 32-degree-of-freedom series-parallel hybrid design. The disparity between the human musculoskele- tal system and the robot’s structure and constraints highlights the complexity of replicating anthropomorphic upper-body motions.

Our approach leverages reinforcement learning (RL) with GPU acceleration to train robust control policies. We employed the Proximal Policy Optimization (PPO) algorithm for its efficacy in stable policy training. The training process incorporated multidimensional reward functions emphasizing motion accuracy, stability, and efficiency, enabling the RH5 robot to reliably perform chin-up motions.

Experimental results demonstrate that the RL-driven method successfully generates chin-up motions that mimic human behavior. However, the results also reveal large saturation of RH5’s joint torques and velocities limitations, underscoring the need for further optimization in its design.

Future work will explore RL-based co-design, integrating simulation and experimental insights to develop an enhanced RH5 design capable of dynamically feasible and versatile performance for highly dynamical tasks.

Keywords: AI-Enabled Robotics, Robot Safety, Computer Architecture for Robotic and Automation
Abstract: Rapid prototyping of robotic controllers trained in domains different from their deployment environments presents numerous challenges, particularly regarding safety concerns. Typical laboratory test bench setups implement different forms of filters or control limitations to meet safety requirements. However, these are not always represented during the training phase, as their inclusion can adversely affect the learning process. To address these challenges, we developed a toolchain consisting of several ROS modules (Fig. 1). We evaluated its transferability from simulation to reality by testing different safety configurations during training. Our experiments explore how incorporating various safety constraints in simulation affects learning efficiency and the robustness of controllers when transferred to real-world applications. Our findings indicate that while including safety constraints during simulation does not necessarily slow down convergence or reduce peak performance, it significantly enhances the transferability of controllers to real-world scenarios where safety measures are mandatory. Conversely, omitting safety constraints during training and applying corrective mechanisms only at deployment can also yield effective outcomes.

Keywords: Robotics in Hazardous Fields, Perception for Grasping and Manipulation, Robotics and Automation in Construction
Abstract: This video demonstrates the capabilities of the autonomous excavator ALICE performing complex tasks during various field tests. ALICE has been developed with focus on applications in decontamination and recovery, where autonomous heavy machinery can work in hazardous areas so that humans no longer need to enter the danger zone. Relevant scenarios include the remediation of landfill sites and the recovery of hazardous materials following a transportation accident. Therefore, the recovery of barrels potentially containing dangerous materials has been demonstrated. Additionally, a field test on a landfill site has been conducted where excavation for reprofiling the terrain as well as loading a dump truck have been performed.

ALICE (Autonomous Large Intelligent Crawler Excavator) is a crawler excavator of the 24 metric ton class. It is equipped with a tiltrotator, a quickcoupler, and various attachments including a sorting grab and several buckets. The excavator has been automated by means of joint angle sensors and a CAN bus interface for digital actuation of the hydraulic joints. For environment perception, LiDAR sensors and cameras provide a 360 degree field of view. The 'Algorithm Toolbox' software developed by Fraunhofer IOSB provides robotic capabilities including localization, mapping, obstacle detection, autonomous driving, autonomous manipulation, and earthmoving.

Keywords: AI-Enabled Robotics, Marine Robotics, Multi-Robot Systems
Abstract: Marine debris poses an increasingly alarming threat to aquatic ecosystems. Conventional methods of sea and ocean cleaning rely heavily on manual collection, a process that has repeatedly demonstrated its inefficiency and extensive demand for resources. This video presents the SeaClear system, a novel multi-robot platform designed to autonomously detect and collect marine debris, thereby offering a more efficient solution to this pressing environmental challenge. An overview of the system is presented, followed by a demonstration of each robot’s capabilities. Leveraging artificial intelligence, the system employs advanced computer vision algorithms for the detection of underwater litter, addressing challenging visibility conditions and hydrodynamic disturbances of underwater environments. Additionally, the video demonstrates navigation and control methodologies to create a map of the identified litter objects and collect those items. The performance of the designed system is shown via field tests conducted in real-world underwater environments, with the video presenting the final demonstration in the region of Dubrovnik, Croatia.

Keywords: Legged Robots, Machine Learning for Robot Control, Optimization and Optimal Control
Abstract: Legged robots are able to navigate complex terrains by continuously interacting with the environment through careful selection of contact sequences and timings. However, the combinatorial nature behind contact planning hinders the applicability of such optimization problems on hardware. In this work, we present a novel approach that optimizes gait sequences and respective timings for legged robots in the context of optimization-based controllers through the use of sampling-based methods and supervised learning techniques. We propose to bootstrap the search by learning an optimal value function in order to speed-up the gait planning procedure making it applicable in real-time. To validate our proposed method, we showcase its performance both in simulation and on hardware using a 22 kg electric quadruped robot. The method is assessed on different terrains, under external perturbations, and in comparison to a standard control approach where the gait sequence is fixed a priori.

Keywords: Human-Robot Collaboration, Human-Centered Robotics, AI-Enabled Robotics
Abstract: Successful adoption of industrial robots will strongly depend on their ability to safely and efficiently operate in human environments, engage in natural communication, understand their users, and express intentions intuitively while avoiding unnecessary distractions. To achieve this advanced level of Human-Robot Interaction (HRI), robots need to acquire and incorporate knowledge of their users' tasks and environment and adopt multimodal communication approaches with expressive cues that combine speech, movement, gazes, and other modalities. This paper presents several methods to design, enhance, and evaluate expressive HRI systems for non-humanoid industrial robots. We present the concept of a small anthropomorphic robot communicating as a proxy for its non-humanoid host, such as a forklift. We developed a multimodal and LLM-enhanced communication framework for this robot and evaluated it in several lab experiments, using gaze tracking and motion capture to quantify how users perceive the robot and measure the task progress.

Keywords: Human and Humanoid Motion Analysis and Synthesis, Rehabilitation Robotics, Motion and Path Planning
Abstract: This paper presents a novel geodesic-based trajectory generation approach for exoskeleton-aided rehabilitation, aiming to replicate natural human joint coordination while meeting clinical requirements. Existing algorithms often fail to effectively mimic human motions across diverse tasks, particularly in the presence of redundant kinematic chains. The proposed method incorporates joint-level constraints and energy minimization principles to generate human-like trajectories. Empirical results using upper-limb data demonstrate superior performance compared to traditional models like minimum-jerk and cubic polynomial approaches, offering enhanced alignment with human motions in both joint and task spaces.

Keywords: Autonomous Vehicle Navigation, Motion and Path Planning, Motion Control
Abstract: Current autonomous driving technologies are resource-intensive in their development, and deploying them on public streets remains technologically challenging and financially unsustainable, contributing to the economic strain in Germany's and Europe’s automotive sector. Large Multimodal Foundation Models (LMFM) are emerging as a framework to tackle challenges in robotics due to their advanced context understanding and capturing causalities across different modalities being trained on various tasks. Fine-tuning LMFMs for specific tasks or domains, such as autonomous driving, promises higher performance. However, despite their potential, current state-of-the-art methods struggle to master the complexities of the driving task. They are often resource-intensive and sample-inefficient, lagging behind the efficiency and effectiveness of human learning processes. In this research, we propose a novel approach to mimic human-level learning for driving, by introducing a high-level multi-task fine-tuning curriculum that divides the driving task into four phases, mirroring the structured progression found in human driving school curricula. Our approach will gradually enhance task complexity while reducing the reliance on expert guidance. Being successful in this approach, we will deliver an end-to-end autonomous driving system capable of mapless navigation while adhering to Road Traffic Regulations (StVO). To increase the reasoning and decision-making capabilities of our base LMFM and bridge the gap between simulation and real-world driving, we introduce a novel continuous fine-tuning technique termed online Iterative Reinforcement Driving Learning from Driving Instructor Feedback/Suggestion (RDL-DIF). Leveraging our extensive experience in autonomous vehicle algorithm development across public roads and racing environments, we aim to achieve a significant 92x reduction in the time required to master the real-world driving task, while simultaneously cutting resource requirements by at least 96x compared to existing state-of-the-art approaches. By streamlining resource demands, we seek to pioneer the next generation of autonomous vehicle software, making scalable and profitable autonomous vehicle systems a reality.

Keywords: Imitation Learning, Learning from Demonstration, Deep Learning in Grasping and Manipulation
Abstract: We introduce a novel approach for object-centric imitation learning for robotic manipulation. Our approach, Keypoint Movement Primitive Diffusion (KMPD), combines diffusion policies, keypoint-based visual imitation learning, and movement primitives for data-efficient, interpretable, and robust object-centric manipulation. Building on the interpretability and geometric constraints of previous work, we employ state-of-the-art vision models to extract semantic representations and keypoint dynamics from human demonstration videos. These features guide a transformer-based diffusion model that predicts the weights for low-dimensional movement primitives, which generate smooth trajectories per keypoint. These keypoint trajectories are aggregated to a 6DoF transformation for rigid objects, or used for a Keypoint-based Admittance Controller for downstream robotic manipulation. Our ongoing work aims to validate this approach in simulation with real-world data, followed by physical robot testing on a variety of tasks. This research paves the way for robust, versatile imitation learning at the intersection of computer vision, machine learning, and robotics.

Keywords: Force and Tactile Sensing, Perception for Grasping and Manipulation, Hardware-Software Integration in Robotics
Abstract: Tactile sensors are crucial for many robotic applications, such as dexterous robotic gripping, prosthetics, and surgical robots. However, only a few companies offer this type of sensor. Additionally, these sensors are too expensive and fragile to be used in many applications. We are developing a vibration-based tactile sensor integrated into a robotic hand as a demonstration. We focus on improving machine learning algorithms for texture recognition, object detection, and localization of touch stimuli using body-borne vibrations, creating a more efficient and cost-effective solution for diverse applications such as robotics. We are transferring the technology to a start-up developing robotics hands and grippers.

Keywords: Human-Robot Collaboration
Abstract: Human-robot collaboration (HRC) enables efficient and flexible task execution in shared workspaces through intelligent robotic systems. However, evaluating such systems based on long-term metrics is challenging because traditional user studies only capture short-term interactions with the robot. Virtual commissioning (VC) addresses this issue by enabling evaluation in a simulated environment, allowing for a comprehensive analysis of dynamic interactions across multiple applications. In addition to a robot model, such a simulation also requires models that accurately represent human behavior. Furthermore, since individual humans do not necessarily follow the same behaviors, we propose that such models should include parameters to model individual human characteristics. Based on our previous work, we propose a human model composed of two components. A Markov Decision Process (MDP) represents high-level behaviors such as collaboration, communication, or rest. On the other hand, a Markov model, derived from demonstrations, selects the individual assembly actions.

Keywords: AI-Enabled Robotics, Semantic Scene Understanding, AI-Based Methods
Abstract: Open-world object manipulation requires a comprehensive understanding of physical scenes and user commands to solve complex tasks. Recent advances in vision-language models (VLMs) have demonstrated capabilities in open-world problems, how to utilise them for fine-grained scene understanding and advanced reasoning of vision-language information remains an open challenge. For this purpose, this study explores using pre-trained VLMs to interpret and infer vision-language information and generate keypoint affordance-based robot actions for task execution. Our method enables a fine-grained semantic understanding of the robotic scene and its elements' spatial relation with zero-shot samples. By prompting the pre-trained VLM, advanced reasoning and common-sense knowledge of the VLMs are utilised to predict the latent logic representation and control steps of language tasks. Then, a keypoint affordance representation method is presented to extract optimal manipulation points for the object of interest, where the VLM, as a motion planner, takes keypoint affordances and visual observations to generate text-based robot path descriptions for task execution. Next, high-level text-based robot paths are linked with low-level robot controller to regular the robot behaviours for task execution. Finally, a set of experiments on various manipulation tasks are performed to validate the proposed methods with evaluation metrics.

Keywords: RGB-D Perception, Probability and Statistical Methods, Learning from Demonstration
Abstract: We propose a probabilistic real-time method for estimating kinematic models of articulated objects by tracking human hand motions as a perceptual prior. Our approach leverages the properties of the human hand to address uncertainties arising from occlusions and noise in visual observations by explicitly modeling these uncertainties. Tested on a novel dataset, our method significantly outperforms recent baseline models in accuracy by 195% and 140%, respectively. Furthermore, we demonstrate that our approach enables robots to manipulate small objects safely, achieving a success rate of 100% with only a single demonstration.

Keywords: Force and Tactile Sensing, Perception for Grasping and Manipulation, Deep Learning Methods
Abstract: Estimating compliance is an important skill in many areas, such as agriculture and the biomedical field. Traditional methods are highly specialized for specific use cases and often suffer from high costs and low portability, making them unsuitable for robotic applications. Due to the increased popularity of vision-based tactile sensors, modern solutions are shifting towards utilizing tactile images for compliance estimation. However, these neural network based solutions still suffer insufficient prediction accuracy, utilize additional analytical features and do not employ state-of-the-art techniques. We propose two new model architectures based on Recurrent Convolutional Networks and Transformers. Our proposed models display improved prediction accuracy across multiple metrics and show that analytical input features are unnecessary to achieve state-of-the-art performance.

Keywords: Robotics and Automation in Agriculture and Forestry, Data Sets for Robotic Vision, Deep Learning for Visual Perception
Abstract: RGB-NIR image registration plays an important role in sensor-fusion, image enhancement and off-road au- tonomy. In this work, we evaluate both classical and Deep Learning (DL) based image registration techniques to access their suitability for off-road forestry applications. NeMAR trained under 6 different configurations, demonstrates partial success however, its GAN loss instability suggests challenges in preserving geomertic consistency. MURF, when tested on off-road forestry data shows promising large scale feature alignment during shared information extraction but struggles with fine details in dense vegetation. Even though this is just a preliminary evaluation, our study necessitates further refinements for robust, multi-scale registartion for off-road forest applications

Keywords: Aerial Systems: Applications, AI-Based Methods, AI-Enabled Robotics
Abstract: Nowadays, unmanned aerial vehicles (UAVs) are commonly used in search and rescue scenarios to gather information in the search area. The automatic identification of the person searched for in aerial footage could increase the autonomy of such systems, reduce the search time, and thus increase the missed person’s chances of survival. In this paper, we present a summary of our novel approach to perform semantically conditioned open vocabulary object tracking that is specifically designed to cope with the limitations of UAV hardware. Our approach has several advantages: (i) it can run with verbal descriptions of the missing person (e.g., the color of the shirt), (ii) it does not require dedicated training to execute the mission, and (iii) can efficiently track a potentially moving person. Our experimental results demonstrate the versatility and efficacy of our approach. Our source code is available on GitHub.

Keywords: Learning from Demonstration, Perception-Action Coupling, Compliance and Impedance Control
Abstract: We present a novel LfD framework for contact-rich manipulation tasks. The core of our approach is the use of environmental constraint as the representation to replicate human demonstration motions. This design choice is based on the belief that environmental constraints serve as an appropriate abstraction that separates general information from the information specific to instances. This separation ensures the acquired manipulation policies generalize to various scenarios that share the same environmental constraints. Following this insight, we design a component that autonomously recognizes environmental constraints in a single demonstration, sensorimotor primitives that deliberately exploit multi-modal feedback provided by those constraints, and a correction module that integrates human corrections to recover from manipulation failures. Extensive real-world experiments, spanning six distinct manipulation tasks, validate the effectiveness of our LfD framework.

Keywords: Wearable Robotics, Rehabilitation Robotics, Prosthetics and Exoskeletons
Abstract: Hybrid exoskeleton combines robotic assistance with functional electrical stimulation (FES) to address the limitations of their respective assistive techniques in the neuro-rehabilitation of motor deficits. The exoskeleton provides external limb support but does not directly activate the neuromuscular system, while FES actively engages the neuromuscular system but faces challenges in providing accurate functional assistance due to the nonlinear, user- and state-dependent behavior of the neuromuscular system in response to electrical stimulation. Resolving redundancy between FES and exoskeleton involves many dimensions including but not limited to nonlinearity, different actuators' dynamics and constraints, and rehabilitation uptake. This paper, therefore, aims to introduce a novel cooperative control approach based on adaptive control allocation that addresses actuator redundancy between FES and exoskeleton while accounting for their limitations.

Keywords: Robot Safety, Reinforcement Learning, Machine Learning for Robot Control
Abstract: Reinforcement learning (RL) controllers are flexible and performant but rarely guarantee safety. Safety filters impart hard safety guarantees to RL controllers while maintaining flexibility. However, safety filters cause undesired behaviours due to the separation of the controller and the safety filter, degrading performance and robustness. This extended abstract unifies two complementary approaches aimed at improving the integration between the safety filter and the RL controller. The first extends the objective horizon of a safety filter to minimize corrections over a longer horizon. The second incorporates safety filters into the training of RL controllers, improving sample efficiency and policy performance. Together, these methods improve the training and deployment of RL controllers while guaranteeing safety.

Keywords: AI-Based Methods, Object Detection, Segmentation and Categorization
Abstract: Object detection is a crucial task in many mobile robotics applications but is also very power hungry. We present an FPGA-based accelerator for YOLOv6, trained on the VOC2007 dataset, using FINN and Brevitas with 6-bit quantization for weights and activations. Key innovations include custom transformations for the SPPF block, optimized fork/join handling, and an shift-and-add replacement for MultiThreshold-ReLU, reducing Look-Up Table (LUT) and Flip-Flop (FF)usage by over 5× and 4×, respectively. The accelerator was successfully deployed on a mobile robot using a cost-effective ZCU102 FPGA, demonstrating its practical viability for real-time autonomous driving.

Keywords: Acceptability and Trust, Human Factors and Human-in-the-Loop, Human-Robot Collaboration
Abstract: Industrial robots are ubiquitous in today's production systems. Here, strict occupational health and safety regulations apply to their operation. As a rule, robots work spatially separated from humans. Only special designed robots are currently allowed to work directly with humans at workplaces in a human-robot collaboration (HRC) setup. HRC can enable companies to meet the challenges posed by the growing complexity of production processes and the need for flexibilization with the aim of increasing efficiency. To achieve these goals, however, humans and robots must be able to work together trustworthy and without fear. Although robots capable for HRC are subject to limitations in terms of allowed speed and load capacity to ensure safe operation, it is not yet known what level of fear and trust enables efficient HRC for the human worker. Here, a transparent interaction design enabling the prediction of robot's movement and intentions can help to ensure a smooth and seamless HRC. Despite many different approaches to achieve transparency in HRC, it is still unclear which transparency mechanisms are best suited for different types of robots and situations, especially when humans have different levels of attention towards the robot. The overall goal of our work is to contribute towards more comprehensible and trustworthy HRC interaction. This leads to a more flexible and human-centered implementation of robots in production. In this paper, we’ll give a brief overview of relevant literature and a comprehensive outlook on our future research.

Keywords: Medical Robots and Systems, Human-Centered Robotics, Physical Human-Robot Interaction
Abstract: Pediatric forearm fractures are common, necessitating diagnostic methods that minimize radiation exposure. Ultrasound offers a safer alternative to X-rays, but conventional imaging is highly operator-dependent and often painful due to direct contact with the injury. This study introduces a robot-assisted, contactless ultrasound system using the UR5 robotic manipulator to enhance imaging accuracy and comfort. We present the system's current development, experimental validation of underwater ultrasound imaging, and motion planning strategies for robot-assisted scanning. Future work will focus on automating image acquisition and integrating 3D tomographic reconstruction to enhance diagnostic precision.

Keywords: Disassembly, Computer Vision for Automation, Industrial Robots
Abstract: Automated disassembly is essential for sustainable end-of-life solutions, promoting a circular economy. This contribution presents an approach to robotic disassembly focusing on CAD-based model extraction, disassembly planning, and state sensing. We introduce a method to generate disassembly sequences from polygon meshes, incorporating non-geometric joints like adhesives. Geometric joints are detected through contact analysis, forming a constraint graph used to determine optimal sequences. Beyond non-destructive disassembly, we propose an automated approach for deriving destructive actions based on symbolic disassembly representations, validated both virtually and physically using a robotic manipulator. To handle deviations between digital and physical models, we compare simulated point clouds with real-time 3D camera data, integrating part movability for accuracy. This adaptive sensing ensures robustness in execution. By combining sequence generation, destructive planning, and real-time sensing, our work advances fully automated, adaptive robotic disassembly for complex products.

Keywords: Medical Robots and Systems, Human-Centered Robotics, Physical Human-Robot Interaction
Abstract: Forearm fractures in children often require X-ray diagnostics, which expose sensitive tissues to ionizing radiation. To reduce radiation exposure, ultrasound offers a radiation-free alternative but is hindered by the need for specialized training and discomfort from probe pressure. The SonoBox addresses these challenges with a robot-assisted, contactless ultrasound method. A robotic manipulator guides the ultrasound probe while the forearm is submerged in a water tank, enabling tomographic imaging without direct pressure.

Keywords: Prosthetics and Exoskeletons, AI-Enabled Robotics, Incremental Learning
Abstract: We propose a minimalist ridge‑regression approach to interpret surface electromyography (sEMG) signals for real‑time ankle prosthesis control. This proof‑of‑concept system implements three open‑loop control modes —position, velocity, and torque— requiring minimal calibration. Three able‑bodied participants completed a Target Achievement Control (TAC) test under varying difficulty. Initial results demonstrate control and task-dependent user performance and highlight distinct trade‑offs between overshoot, throughput, and user effort, which were found to be minimized for velocity control.

Keywords: Medical Robots and Systems, Human-Centered Robotics, Natural Dialog for HRI
Abstract: This work introduces a novel concept for a robotic system that scans its environment with a camera to detect the presence of humans, interacts with them, and subsequently gathers their vital signs contactless. Upon detecting a person, the system promptly acquires key health parameters — such as heart and respiratory rate or body temperature — by autonomously orienting the camera mounted at the end of the robotic arm in line-of-sight with the face of the detected person. Afterward, the acquired vital data is securely stored. By combining event-driven detection with unobtrusive monitoring, the proposed solution offers a new robot-based approach for routine contactless vital sign collection, potentially reducing the burden on both medical staff and patients. Building on this conceptual framework, the next phase involves developing a fully functional prototype.

Keywords: Human-Centered Robotics, Telerobotics and Teleoperation, Mobile Manipulation
Abstract: Intuitive teleoperation interfaces are an essential part of efficient data collection for robot learning. A high sense of embodiment, low physical- and cognitive workloads are ideal for a pleasant user experience and good performance. To reliably assess those goals, we conducted an extensive user study (N=20) to evaluate 4 different teleoperation and feedback interfaces on 5 diverse tasks that comprise an extended task sequence in a complex manipulation scenario. Preliminary results indicate that the usage of a virtual reality interface increases task completion time and workload on the user. Furthermore, our whole body controller is more physically demanding, but leads to less frustration while using the VR, compared to controlling base and arm separately.

Keywords: Agricultural Automation, Autonomous Vehicle Navigation, Motion and Path Planning
Abstract: In this paper we present a PhD research proposal in the domain of agricultural robotics. Utilizing a 4WISD mobile robot platform and a 6-DoF manipulator, we aim to tackle multiple challenges in both mobile robot control and navigation as well as robot-plant interaction.

Keywords: Probabilistic Inference, Intention Recognition, Motion and Path Planning
Abstract: Goal inference is crucial in robotics, enabling effective collaboration in human robot interaction (HRI) and assisted teleoperation. Current approaches often rely on Markov decision processes (MDP) and maximum entropy principles to infer intentions by integrating over trajectory space. However, these methods commonly employ local approximations around optimal trajectories, which oversimplify the integration and result in unimodal trajectory predictions. They predominantly consider straight-line paths or user input-related costs, neglecting geometric and contextual constraints such as obstacles. This paper proposes a Geometrically-aware goal inference framework that integrates motion planning with Bayesian inference. By leveraging motion planning as inference to generate multimodal trajectory distributions and employing belief updates through Sequential Monte Carlo methods, our approach demonstrates the efficiency of capturing goal-directed behavior in complex environments. This preliminary study highlights the promise of combining motion planning with goal inference and motivates future research toward more comprehensive evaluations.

Keywords: Reactive and Sensor-Based Planning, Behavior-Based Systems, Probabilistic Inference
Abstract: Robot behavior needs to be robust. To achieve robustness, traditional approaches attempt to anticipate all possible situations the robot could encounter and account for them in a policy. But how do we represent the behavior so that it also remains robust in unanticipated situations? Here, we present AICON (Active InterCONnect), a framework that generates behavior without directly encoding it. Instead, we dynamically compose multiple sensorimotor regularities to simultaneously estimate relevant states and obtain action gradients. Using these gradients, we can generate robust robotic behavior for real-world tasks even for unanticipated scenarios and large disturbances. We also show that AICON can be used to study biology as it possesses characteristics similar to biological information processing.

Keywords: Robotics and Automation in Agriculture and Forestry, Environment Monitoring and Management, Biologically-Inspired Robots
Abstract: Forests provide essential resources and services to humanity, yet monitoring and preserving them remain challenging due to the difficulty of accessing critical areas, particularly dense forest canopies. Traditional methods require biologists to climb trees and manually place sensors, a time-consuming and dangerous process. While robotic solutions, such as swarms of free-flying drones, have demonstrated effective navigation through the canopy for mapping and data collection, they remain limited by energy constraints, noise pollution, and insufficient capabilities for precise physical interactions. Alternative aerial sensor deployment strategies, including bio-gliders and sensor shooting techniques, offer promising solutions but often lack efficient data retrieval methods, necessitating human intervention. To address these limitations, we propose an aerial tethered perching and disentangling mechanism designed for long-term environmental monitoring. The system integrates a ring mechanism on the drone with a suspended pod featuring a winding system and propellers, enhancing mobility, stability, and energy efficiency. By perching on tree branches, the drone can shut down its motors, significantly reducing energy consumption and noise, thereby minimizing ecological disturbances. This capability enables data collection for extended periods without frequent battery replacements, making the system particularly well-suited for long-term ecological studies and biodiversity research.

Keywords: Force and Tactile Sensing, Probabilistic Inference, Deep Learning Methods
Abstract: Hardness is a key tactile property perceived by humans and robots. In this work, we investigate information-theoretic active sampling for efficient hardness classification using vision-based tactile sensors. We assess three probabilistic classifiers and two uncertainty-based sampling strategies on a robotic setup and a human-collected dataset. Results show that uncertainty-driven sampling outperforms random sampling in accuracy and stability. While human participants achieve 48.00% accuracy, our best method reaches 88.78% on the same objects, highlighting the effectiveness of vision-based tactile sensors for hardness classification.

Keywords: Robot Safety, AI-Enabled Robotics
Abstract: Ensuring safe robot interactions in human environments requires adhering to common-sense safety (e.g., preventing spilling of water by keeping a cup straight). While safety in robotics is extensively studied, semantic understanding is rarely considered. We propose a semantic safety filter that certifies robot actions against semantically defined and geometric constraints. Our approach builds a 3D semantic map from perception inputs and uses large language models to infer unsafe conditions, which are enforced via control barrier certification. We validate our framework in teleoperated and learned manipulation tasks, demonstrating its effectiveness in real-world scenarios beyond traditional collision avoidance.

Keywords: Motion and Path Planning, Multi-Robot Systems, Aerial Systems: Mechanics and Control
Abstract: We present db-ECBS, a novel kinodynamic motion planner for multi-robot teams that can directly reason about interaction forces that occur for example in aerodynamic teams that fly in close proximity to each other. Algorithmically, we generalize db-CBS to include non-admissible heuristics that guide the search to avoid conflicts and we augment the state space during the search and optimization to include aerodynamic forces. Our approach is probabilistically complete, asymptotically bounded suboptimal, and unlike differentially-flatness-based planners can directly reason about actuation constraints. Empirically, we demonstrate that db-ECBS can produce trajectories that are less than half the cost of existing planners and that the interaction-awareness is in particular important for very dense scenarios.

Keywords: Computer Vision for Automation, Data Sets for Robotic Vision, Human Factors and Human-in-the-Loop
Abstract: Accurate intrinsic and extrinsic camera calibration can be an important prerequisite for AI-powered robotic applications that rely on vision as input. While there is ongoing research on enabling calibration using natural images as input, many systems in practice still rely on using designated calibration targets with e.g. checkerboard patterns or April tag grids. Once calibration images from different perspectives have been acquired and feature descriptors detected, those are typically used in an optimization process to minimize the geometric reprojection error. For this optimization to converge, input images need to be of sufficient quality and particularly sharpness; they should neither contain motion blur nor rolling-shutter artifacts that can arise when the calibration board was not static during image capture.

In this short report, we present a novel calibration image acquisition technique controlled via voice commands recorded with a clip-on microphone, that can be more robust and user-friendly than e.g. triggering capture with a remote control, or filtering out blurry frames from a video sequence in postprocessing. To achieve this, we first align audio and video recordings by a hand clapping gesture, and then use a state-of-the-art speech-to-text transcription model with accurate per-word timestamping to capture trigger words with precise temporal alignment.

Keywords: Human-Robot Collaboration, Industrial Robots, Learning Categories and Concepts
Abstract: The increasing integration of collaborative robots (cobots) into industrial settings requires intuitive methods for programming and skill acquisition, as these cobots work alongside humans on assembly lines. Traditional cobot programming requires specialized expertise, creating bottlenecks in adapting them to dynamic industry environments. This paper explores conversational programming as an alternative, enabling workers without programming skills to interact with cobots using natural language. Using a 2.5D structure building task that is designed to simulate the industry assembly scenarios, we study the capabilities of large language models (LLMs) for conversational code generation. Our results show the models are able to generate accurately what we call ``first-order code'' (i.e., sequences of instructions) and struggle with producing ``higher-order code'' (i.e., code containing abstractions and loops). Finally, we outline future research directions aimed at enhancing adaptability, abstraction capabilities, and real-time learning efficiency, paving the way for flexible and human-friendly industrial automation.

Keywords: Acceptability and Trust, Big Data in Robotics and Automation, AI-Enabled Robotics
Abstract: The concept of robot has undergone many changes since its initial inception. From the initial description of man-made workers in Karel Čapek's R.U.R., the theme of the interactions between society and its artificial offspring has been central to fiction and, with time, to public debate as well. In this paper, we would like to analyze the various representations of robots in fiction, philosophy and in modern culture, and examine how these representations are connected to humanity's conception of itself, as well as how research, fiction and communication interact with each other.

Keywords: Deep Learning for Visual Perception, Data Sets for Robot Learning, Simulation and Animation
Abstract: The lack of large-scale annotated depth datasets poses a challenge for training machine learning models in robotics and autonomous systems. To address this, we present a method for generating synthetic, semantically labeled depth data using virtual sensors in a 3D simulation environment. This approach supports LiDAR, Sonar, and Time-of-Flight (ToF) cameras, incorporating realistic noise modeling, environmental effects, and customizable sensor parameters. By leveraging procedurally generated landscapes and photogrammetry-based reconstructions, diverse datasets can be created to improve AI-based depth perception. Exemplary applications demonstrate the effectiveness of this synthetic data in enhancing machine learning performance.

Keywords: Software Architecture for Robotic and Automation, Software Tools for Robot Programming, Software-Hardware Integration for Robot Systems
Abstract: No matter whether it is in simulation or in the lab – setting up robotics experiments is challenging, especially if they target both types of deployments. While a plethora of different robotics software frameworks exist, they usually make a trade-off between complexity and versatility. Drawing from this rich ecosystem, we present Robot Control Stack (RCS), a robotics software stack that is designed to resolve this conflict by combining established robotics tools and frameworks through a lightweight multi-layered API that unifies simulation and real-world robot control. At the core of RCS is a minimalist architecture built around the concept of Gymnasium environment wrappers that can be easily extended at different levels of abstraction. In this paper, we motivate and outline the design of RCS and its architecture, report on the current state of development and initial use cases, and highlight how it supports the training and evaluation of robotics foundation models.

Keywords: Imitation Learning, Representation Learning, Multi-Modal Perception for HRI
Abstract: Imitation Learning (IL) has shown strong performance in teaching robots complex tasks but often struggles with noisy or incomplete sensory inputs. Multi-sensor setups are prone to failures, dropouts, and noise, requiring robust multimodal integration. This paper presents Robust Multiview- Multimodal Diffusion Policies (RM-MDP). RM-MDP improves robustness and interpretability through multi-view disentanglement and orthogonality constraints in IL. Evaluations on the Libero and Colosseum benchmarks demonstrate its superior performance compared to baseline methods.

Keywords: Surgical Robotics: Laparoscopy, Imitation Learning, Learning from Demonstration
Abstract: Imitation learning (IL) methods have become popular in the automation of surgical tasks. We explore the effectiveness of IL in enabling a robot to perform a thread-in-hole task, a common training exercise for laparoscopic surgeons. In this setup, visual feedback is provided by a single stereoscopic camera, whereas previous studies usually employ multiple cameras to ensure robustness and generalization. Thread-in-hole is both simple enough to be learned with stereoscopic vision while also serving as a foundation for more complex surgical procedures, such as threading or performing precise sutures. We observe that our learned policy solves the task with a success rate of 43%, demonstrating its potential for improvement and further refinement.

Keywords: Simulation and Animation, Reinforcement Learning, Dexterous Manipulation
Abstract: Training robot policies in simulation is becoming increasingly popular; nevertheless, a precise, reliable, and easy-to-use tactile simulator for contact-rich manipulation tasks is still missing. To close this gap, we develop TacEx -- a modular tactile simulation framework. We embed a state-of-the-art soft-body simulator for contacts named GIPC and vision-based tactile simulators Taxim and FOTS into Isaac Sim to achieve robust and plausible simulation of the visuotactile sensor GelSight Mini. We implement several Isaac Lab environments for Reinforcement Learning (RL) leveraging our TacEx simulation, including object pushing, lifting, and pole balancing. We validate that the simulation is stable and that the high-dimensional observations, such as the gel deformation and the RGB images from the GelSight camera, can be used for training. The code, videos, and additional results will be released online url{https://sites.google.com/view/tacex}.

Keywords: Gesture, Posture and Facial Expressions, Motion Control
Abstract: Accurate monitoring of finger flexion is essential for advances in hand gesture recognition. Existing studies focus on classifying the gesture in its totality, without considering the flexion of individual fingers, since determining the degree of finger flexion may lead to the classification of unclassified gestures. This study employs a multimodal sensing approach that integrates direct and indirect measurement techniques. A strain sensor based on TPU/CNT is proposed for direct quantification of finger flexion. At the same time, eight nanocomposite pressure sensors are strategically positioned to indirectly capture variations in forearm muscle activity associated with finger movement. During finger flexion, the strain sensor shows an increase in impedance of approximately 0.5 × 105 Ω, while pressure sensors exhibit a smaller decrease, reflecting localized muscle activation. In the extended position, the strain sensor maintains a static impedance of 8 × 105 Ω, and the pressure sensors remain within 4×105 Ω to 6×105 Ω. This dual-sensing method enables the future classification of unclassified gestures, which is essential for improving hand gesture classification, human-machine interaction, and robotic control.

Keywords: Whole-Body Motion Planning and Control, Telerobotics and Teleoperation, Mobile Manipulation
Abstract: We present a novel approach for whole-body teleoperation of mobile manipulation robots, focusing on complex household tasks that require coordination of arms manipulation and robot-body posture while enabling multiple contact points along the robot’s kinematic chain. Current state-of-the-art methods often overlook the synergy between different parts of the robot’s body, particularly in humanoid robots, as they typically rely on end-effector actions alone. In contrast, we propose a teleoperation framework that supports multi-embodiment and multi-contact teleoperation, allowing for efficient whole-body control that enhances task execution in real-world scenarios and brings robots closer to human-like behaviors. Videos of early real-robot experiments can be found on https://sites.google.com/view/aawbt/home.

Keywords: Computer Vision for Medical Robotics, Surgical Robotics: Planning, RGB-D Perception
Abstract: Object interactions in robotics and computer vision often rely on explicit registration techniques, which can be computationally expensive. Additionally, these methods often suffer from initialization errors, particularly for deformable objects. In this work, we present neural occupancy functions as an alternative to registration-based approaches for deformable object understanding. We assume the object of interest is known prior. Our approach allows accurate understanding of deformable objects with a low latency (less than 30 ms).

Keywords: Big Data in Robotics and Automation, Deep Learning in Grasping and Manipulation
Abstract: Large vision-language-action (VLA) models have shown remarkable capabilities for learning general robot policies. However, the predominance of English in both the large language model (LLM) backbone training data and the robotics data limits the accessibility of these models. Specifically, training such policies for other languages, such as German, extends their usefulness to the non-English-speaking rest of the world. We present AllmAN, the first German VLA model, built upon LeoLM — a Llama 2-based LLM specifically fine-tuned on large German datasets. To train AllmAN, we machine- translate several English vision-language and VLA datasets into German. We then adapt the Prismatic and OpenVLA training pipelines to create our German VLA model. Through comparative analyses with OpenVLA, we demonstrate the importance of incorporating German language capabilities within the base model. Our findings underscore the importance of training VLAs in other languages beyond English. This work serves as a proof-of-concept for multi-language VLAs, paving the way for broader, more inclusive robotics applications.

Keywords: Actuation and Joint Mechanisms, Humanoid and Bipedal Locomotion, Passive Walking
Abstract: The dynamic and efficient nature of human gait is ever elusive to robots, despite the impressive progress of humanoids over the past years. Two of the driving factors of the growth have been reduced cost of specialty hardware and improved modeling and control approaches. While these factors have facilitated a proliferation of impressive humanoids, there is a plateau of performance which will be reached with the current hardware solutions. To take the next step towards bio-mimetic gaits and dynamic motion, new transparent actuation concepts will be required. To this end, a direct-drive Axial-Flux Permanent- Magnet (AFPM) actuator is being developed for use in the hip joint of the humanoid robot LOLA at the Chair of Applied Mechanics. Using parametric simulation and optimization methods while considering various performance aspects and manufacturing constraints, a three-phase AC YASA AFPM actuator prototype with a predicted peak torque of 284 Nm at an RMS phase current of 14 A has been developed. Testing of the manufactured prototype has resulted in a performance not matching the predictions. This can be attributed to three aspects: 1. Optimistic torque prediction techniques 2. Mechanical design deficiencies leading to deformations and stator-rotor contact 3. Thermal management short-comings. The focus of this work is the analysis potential thermal management techniques for the motor. With the goal of minimizing system complexity while enabling maximal increase of short-term peak-torque, the use of passive phase-change-materials has been determined to be effective.

Keywords: Soft Robot Applications, Grippers and Other End-Effectors, Robotics and Automation in Agriculture and Forestry
Abstract: This research project focuses on developing adaptive soft robotic grippers tailored for multicropping in sustainable agricultural systems such as polycultures. Traditional robotic grippers, mainly designed for controlled environments, struggle to meet the diverse demands of harvesting in real agricultural scenarios with multicropping. The research aims to address these limitations by designing grippers that utilize multi-functional materials and structures to automatically adjust stiffness and shape, inspired by biological systems. By embedding proprioceptive and exteroceptive sensing, the grippers will enhance their “embodied intelligence” to interact with varied crops in unstructured environments. This approach seeks to replace centralized control with adaptive, inherent design features, enabling real-time responsiveness. The research includes field testing, aiming to validate these robotic systems’ practical viability in real settings and thereby support more sustainable agricultural practices.

Keywords: Bioinspired Robot Learning, Biomimetics, Biologically-Inspired Robots
Abstract: Achieving human-like motor performance in bipedal muscle-actuated systems remains difficult given the complex control dynamics and over-actuation. To address this, we harness the embodied intelligence inherent in musculoskeletal systems, drawing inspiration from the natural growth process. While optimizing motor tasks in simulation, we implement abrupt morphological transitions— from a 4-year-old to a 12-year-old, and finally to an adult. Two developmental strategies are explored: one mimicking human ontogenetic growth and another with even scaling of body segments. Our results reveal that both approaches surpass the nondevelopmental option, with bioinspired morphological development holding particular promise for enhancing efficiency and robustness. This advancement brings us closer to achieving human-like performance, ultimately bridging the gap between biological and robotic learning.

Keywords: Human-Robot Collaboration, Behavior-Based Systems, Multi-Modal Perception for HRI
Abstract: The CoMeSy project is developing a system for multimodal interaction between humans and cobots, where the cobot acts as an intelligent assistant. The system uses speech and gestures as input, and responds with speech, sounds, actions, and visual feedback. A key challenge is dynamically creating action plans based on human input, world knowledge, and visual perception. The system integrates several technologies, including speech recognition and synthesis, image processing, object detection, hand tracking, and acoustic feedback. Currently in development, the project aims to address intelligent communication, situational understanding, dynamic planning, reactive behavior, and robust handling of interruptions, with plans for empirical evaluation.

Keywords: Model Learning for Control, Motion and Path Planning, Reinforcement Learning
Abstract: Integrating learning methods into classic control approaches has gained attention in recent years. The demand for faster yet more accurate control of increasingly complex systems motivates the development of novel techniques. However, comprehensive analyses of the various research directions and direct comparisons between different control methodologies on the same robotic systems considering real-time capability of the controller and unknown environments are still limited. This work aims to bridge the gap between the AI and control systems communities by providing an extensive review on incorporating learning methods into traditional MPC. Building upon these theoretical foundations, a simulation environment was created to systematically compare RL, MPC, and Learning-Based MPC (LB-MPC) schemes. Three mechanical test cases of increasing complexity were employed: a cart-pole system, a cart-double-pole system, and a jumping humanoid robot. The methodologies were evaluated based on several metrics, including computational efficiency, adaptability to changing environmental conditions, and robustness against unmodeled dynamics and uncertainties. The results demonstrate that integrating learning techniques into MPC significantly enhances performance in the presence of imperfect modeling. Conversely, incorporating MPC principles into traditional RL schemes substantially improves their sample efficiency.

Keywords: AI-Enabled Robotics, Cognitive Control Architectures, Task Planning
Abstract: Autonomous cognitive robots excel at routine tasks but struggle with novel situations that require adaptability and problem-solving. General-Purpose Service Robots (GPSRs) must adjust to diverse tasks and dynamic environments, but programming for all possible scenarios is labor-intensive and often results in inflexible behavior. Generated control sequences frequently face execution failures and incomplete information, requiring robust verification mechanisms.

This work integrates Large Language Models (LLMs) with the Cognitive Robot Abstract Machine (CRAM) to generate and verify adaptive control sequences. Using Retrieval-Augmented Generation (RAG), LLM agents create action plans represented as a Flanagan Model, which undergoes semantic verification through mental simulation in a digital twin environment. The Plan Executive, running on robot platforms with ROS middleware, ensures executable behaviors by managing spatial constraints, collision checks, and failure recovery.

Experiments show that LLMs, combined with domain knowledge and mental simulation, can generate reliable and executable robot actions without explicit API specifications. The digital twin validation process proves essential for identifying errors and refining robot behaviors. Future work will focus on improving offline models, failure-handling feedback, and optimizing generative AI-driven autonomy in dynamic environments.

Keywords: Robot Safety, Big Data in Robotics and Automation
Abstract: Robot foundation models hold the potential for deployment across diverse environments, from industrial applications to household tasks. While current research focuses primarily on the policies' generalization capabilities across a variety of tasks, it fails to address safety, a critical requirement for deployment on real-world systems. In this paper, we introduce a safety layer designed to constrain the action space of any generalist policy appropriately. Our approach uses ATACOM, a safe reinforcement learning algorithm that creates a safe action space and, therefore, ensures safe state transitions. By extending ATACOM to generalist policies, our method facilitates their deployment in safety-critical scenarios without requiring any specific safety fine-tuning. We demonstrate the effectiveness of this safety layer in an air hockey environment, where it prevents a puck-hitting agent from colliding with its surroundings, a failure observed in generalist policies.

Keywords: Optimization and Optimal Control, Human-Aware Motion Planning, Autonomous Vehicle Navigation
Abstract: Safe and efficient navigation in crowded environments remains a critical challenge for robots. Classical methods decouple human motion prediction from robot motion planning, which neglects the closed-loop interactions between humans and robots. This can lead to inefficiencies such as the Freezing Robot Problem (FRP), where neglecting to account for interaction prevents a robot from finding a plan. We present Safe and Interactive Crowd Navigation (SICNav), a bilevel Model Predictive Control (MPC) framework that combines prediction and planning into one interactive optimization problem. SICNav explicitly incorporates Optimal Reciprocal Collision Avoidance (ORCA)-modeled human predictions as a constraint in the robot’s motion planning MPC optimization problem, enabling the robot to produce collision-free paths that account for interactions. Extensive simulations and real-world experiments validate SICNav’s ability to achieve safe, efficient, and interactive navigation. This extended abstract summarizes work published in IEEE Transactions on Robotics [1].

Keywords: Robot Safety, Collision Avoidance, Engineering for Robotic Systems
Abstract: Assuring the safety of autonomous mobile robots is a core challenge due to the situational complexity and handling of corner cases. Consequently, robotic operating capabilities must be restricted to prevent hazardous situations. Unfortunately, this strategy can lead to an unacceptable performance decrease since a robot’s autonomy could theoretically handle a situation safely, but providing guarantees is impossible. Therefore, this paper proposes an approach for assessing risks during runtime based on situational awareness. This framework leads to a dynamic adjustment of safety measures and avoids unnecessary safety restrictions.

Keywords: Field Robots, Contact Modeling, Soft Robot Applications
Abstract: This paper examines the intersection of robotics and geomechanics, focusing on how robots interact with soft terrains. In particular, we present this new area of robotics research as a critical component of achieving robust and flexible robots operating in natural, typically soft, and unstable terrains. A new paradigm, combining state-of-the-art geomechanics and robotics, is proposed, highlighting some of the challenges that we are facing, as well as presenting a very short literature review of some of the work that has been carried out to date. The paper ends with some research challenges, opportunities ahead, and conclusions.

Keywords: Legged Robots, Simulation and Animation, Field Robots
Abstract: This extended abstract introduces a simulation environment capable of dynamic computation for realistic deformable terrain with quadruped robot interaction. Quadruped robots are capable of mobility on varied terrains. This includes deformable terrains such as soft and clay soil. When a controller is developed based on virtual rigid terrain, its performance may not be good enough in deformable terrains. Thus, this paper focuses on an approach to connect two different frameworks, one for developing a robot controller in a sophisticated manner and the other for testing its performance in a simulation environment with computationally less expensive and more accurate deformable terrain.

Keywords: Legged Robots, Simulation and Animation, Learning from Experience
Abstract: Animals exhibit an extraordinary capability for agile locomotion across various natural terrains. However, replicating this adaptability by developing controllers capable of traversing on soft and deformable terrains, especially at high speed, remains a challenge. Our study proposes an end-to-end learned approach for a fast and robust controller that aims to guide quadruped robots through a wide range of deformable substrates such as soft soil, loose gravel, rough, rocky, wet, muddy, etc. We use a simulation environment that can be parameterized to represent diverse types of terrain that closely mimic the natural world and train a policy network via model-free reinforcement learning. Our proposed approach includes the key components (1) implementing a multiscale granular media model for real-time simulation of terrain dynamics that combines an active zone resolved as distinct particles on top of a resting compliant terrain, (2) using a terrain adaptive curriculum for command velocities, and (3) an online adaptation module strategy that can implicitly identify the terrain properties for easy sim-to-real transfer to physical robot leveraged from prior works.

Keywords: Legged Robots, Contact Modeling, Simulation and Animation
Abstract: Accurate modeling of foot-terrain interactions is critical for enabling stable and efficient locomotion in legged robots, particularly on deformable terrains. This study evaluates three contact models—spring-damper, 3D Resistive Force Theory (3D-RFT), and granular media models using a simulation environment based on a single-leg robot. Unactuated simulations and simulations with control were performed to assess the performance of these models on diverse terrains. Simulations revealed that all three models stabilized at the robot's weight and achieved low tracking errors under proportional-derivative (PD) control. These findings underscore the importance of model selection based on application requirements and provide insights for advancing legged locomotion on diverse terrains.

Keywords: Robust/Adaptive Control, Motion Control, Reinforcement Learning
Abstract: We present a unified framework for high-speed autonomous vehicle control integrating four key advancements in Nonlinear Model Predictive Control (NMPC) for high-dimensional, real-time control systems: (1) Reduced Robustified NMPC enabling real-time computation through sensitivity-based uncertainty propagation, (2) Stochastic NMPC with adaptive Uncertainty Propagation Horizon (UPH) to prevent covariance explosion, (3) Deep Reinforcement Learning (DRL)-driven parameter adaptation for dynamic disturbance handling, and (4) Safe Reinforcement Learning (RL) with Pareto-constrained weight optimization. Validated through extensive simulation and real-world testing at speeds up to 37.5 m/s (135 km/h), our framework achieves real-time performance with 140 Hz solving frequencies, ensuring 98.77% constraint satisfaction and sub-0.1m tracking accuracy. All implementations are available open-source at github.com/bzarr/TUM-CONTROL.

Keywords: Space Robotics and Automation, Motion and Path Planning, Intelligent and Flexible Manufacturing
Abstract: Space exploration relies on increasingly larger telescopes. However next-generation space telescopes may exceed the capacity of current launch vehicles, making in-orbit assembly necessary. To address the challenges associated with on-orbit assembly, the EU-funded project PULSAR has developed and demonstrated technology for the precise robotic assembly of large structures in space In PULSAR, the construction of an 8-meter diameter primary mirror of a space telescope is investigated through software simulations and hardware-in-the-loop demonstrators. One of the demonstrators is dLSAFFE — the Demonstrator of Large Structure Assembly in a Free-Floating Environment. Its core objective is to demonstrate constrained motion planning in the context of the assembly process using representative software and hardware at near one-to-one scale. To lift weight restrictions and to simulate microgravity, the experiments were set in an underwater environment, in the large underwater test basin at DFKI in Bremen. The assembly is performed by a robotic assembly system using modular mirror tiles. The robotic system consists of a manipulator arm by GraalTech with a reach of about two meters and a linear rail as a prismatic joint at the base of the manipulator. With a diameter of approximately 1.3 meters, the size of the mirror tiles is substantial compared to the manipulator arm; fully assembled, the primary mirror even exceeds the reach of the manipulator. To overcome this hurdle, multiple mirror tiles are first assembled to form a large subassembly, which can then be manipulated as a single unit. The video shows the integration of dLSAFFE with the linear rail and the GraalTech manipulator mounted to the central structure of the spacecraft mock-up. The video also shows tests in the large basin at DFKI in Bremen, demonstrating the motion planning and manipulation of a subassembly of mirror tiles.

Keywords: Telerobotics and Teleoperation, Field Robots, Robotics in Hazardous Fields
Abstract: This video presents KRIS, a semi-autonomous robot for material transport in disaster relief situations. The system was developed by the German Research Center for Artificial Intelligence (DFKI) for the German Federal Agency for Technical Relief (THW) to addresses the limitations of manual transport in hazardous environments.

KRIS is equipped with four servomotors each with 250nm torque. The sensor suite is equipped with a LiDAR laser scanner, and an RGB depth camera for precise environment perception. Sensor data processing and SLAM are handled by an Nvidia Jetson Orin AGX, while motor control and low-level operations run on a Jetson Nano and Raspberry Pi. An Xsens IMU ensures accurate motion tracking. The system operates for up to 4 hours on a main battery, with a backup battery enabling seamless swaps.

KRIS addresses the risk of falling during use by developing an easy-to-use assistance robot based on hand gesture control. The operator is recognised by the RGB camera via an Aruco marker. Using a neural network, the system recognises a catalogue of pre-trained body gestures taken from the THW catalogue. The robot recognises its surroundings via LIDAR and can create a map of its environment using the high-level computer. With the help of this map, the robot can navigate safely between two memorised points and, for example, fetch new material during operations.

Keywords: AI-Enabled Robotics
Abstract: The EASE and AICOR Virtual Research Building (VRB) is an online platform designed to support collaborative research and education in everyday activity science and engineering (EASE) and artificial intelligence powered cognition-enabled robotics (AICOR). Using a containerized software ecosystem and scalable cloud infrastructure, the VRB provides virtual laboratories for simulations, research development, reproducibility and education. Researchers can create and share virtual labs with the community, accelerating research workflows, supporting modular robot integration, and enabling rapid validation of algorithms and software without worrying about hardware and software dependencies. The presented regular video highlights how the VRB connects to real-world robotics through photorealistic semantic digital twins and VR applications. It also demonstrates different application fields of virtual labs. For example, by providing an interactive textbook for the education of students and other researchers; and by providing virtual labs to explore the use of plan-based robot control with a variety of robots and environments, to test out robot perception algorithms, or to query robot knowledge services.

Keywords: AI-Enabled Robotics, Computer Vision for Automation, Engineering for Robotic Systems
Abstract: Industries like aerospace, automotive, and manufacturing need strong materials like carbon fibre. They require advanced automation to produce efficiently and maintain quality. Carbon fibre parts can have defects that affect their strength, such as surface flaws, voids, and misalignments during production. Traditional methods for finding these defects, such as manual inspection and basic automation, often take too much time and can lead to mistakes. This research introduces a hybrid robotic system that combines robotic automation with Artificial intelligence (AI)-based Oriented Fast ad Rotated Brief (ORB), and Convolutional Neural Networks (CNNs) to improve defect detection in carbon fibre parts during manufacturing. This system aims to enhance precision and reduce the need for human involvement. The hybrid system consists of a robotic arm programmed using RoboDK, an inspection camera, and a CNN model based on deep learning. The robot moves carbon fibre parts in a controlled setting while the AI approaches analyzes images from the camera to find and identify surface defects like cracks, fibre misalignment, and voids. The model will be trained using images of defective and non-defective carbon fibre parts, which helps it accurately recognize different defects. Using advanced image processing techniques, the system detects defects in real time, improving the efficiency and reliability of quality control. RoboDK simulation software helps model and optimize the robotic arm’s movements and interactions with the parts. This approach makes it easy to pick up and inspect the parts. The simulation also checks how well the system works in different situations, such as with various part angles, lighting, and camera positions. The hybrid robotic system has important benefits. It can quickly detect defects and decide whether to keep or reject a part during manufacturing. Automating this process improves accuracy, reduces labour costs, and makes production more efficient. Combining RoboDK and AI for inspection supports the goal of using advanced robotics and machine learning in industrial manufacturing. This research presents a new way to automatically find defects in carbon fibre parts using robots and AI.

Keywords: Human-Centered Robotics, Human-Robot Collaboration, Datasets for Human Motion
Abstract: Our research focuses on optimizing human-robot interactions (HRI) to enhance efficiency, safety, and satisfaction in occupational settings. We aim to create environments where robots and humans can work harmoniously, adapting robots to human needs rather than vice versa.

In intralogistics, we explore approaches to improve coexistent interactions with mobile robots, such as human-aware navigation and explicit communication tools. Our CoHEXist setup, where humans and a robot interact within a hexagonal space, allows us to track movements and derive metrics like time-to-collision and time taken for interactions. The setup generates 10-25 realistic interactions every 15 minutes.

We also address the prevention of musculoskeletal disorders (MSDs), a common work-related disease. While robots can mitigate MSD risks, they may also introduce new hazards if improperly implemented. Traditional physical load assessments, like the Rapid Upper Limb Assessment (RULA), are often inadequate for work environments with human-robot collaboration. In the Operating Theatre, we use infrared and RGBD-Cameras to track staff, robots and instruments in real-time, capturing 8 hours of footage. Our camera-based system offers high spatial and temporal resolution, enabling automated posture assessment. Currently, we are refining this system for real OR environments and building a database to optimize robotic assistance systems for all stakeholders.

Keywords: Rehabilitation Robotics, Software-Hardware Integration for Robot Systems, Acceptability and Trust
Abstract: Stroke often results in impaired motor function, particularly in upper limb movements. Functional Electrical Stimulation (FES) has demonstrated potential in restoring movement through direct muscle activation. However, its standalone efficacy is constrained by challenges such as rapid muscle fatigue, discomfort, and incomplete motion execution. To address these limitations, hybrid rehabilitation systems integrating FES with robotic exoskeletons offer a promising alternative, combining neuromuscular activation with external mechanical support. The video presentation highlights recent advancements in a hybrid neuroprosthesis for stroke rehabilitation, focusing on hardware innovation, control methodologies, and clinical translation. The research explores the interplay between FES and robotic assistance, with particular attention to adaptive, personalised interaction interfaces and strategies. These research and innovation efforts aim to enhance usability and acceptance among diverse patient populations. By integrating robotics with FES techniques, hybrid neuroprosthetic systems provide a novel and effective approach to motor rehabilitation, paving the way for more personalised functional recovery solutions for individuals with neurological impairments.

Keywords: Human-Robot Collaboration, Haptics and Haptic Interfaces, Force Control
Abstract: The video shows RoboGlove, a hand exoskeleton for teleoperating robots with natural movements while providing haptic feedback, thus merging machine efficiency with human intuition. It opens new possibilities for industrial automation, such as - Operating robots in challenging environments - Performing tactile tasks requiring precision - Simplifying complex processes with intuitive control

The RoboGlove, developed through student projects at Cologne UAS, evolved from controlling mechanical hands to simulating a computer mouse, and later focused on force feedback and robotic arm teleoperation. Inspired by prototypes like HapticGlove and Dexmo, its design was refined through own experience.

Currently, it controls a 2FG7 gripper on a UR5e robotic arm with one finger, using a 3D-printed exoskeleton secured with Velcro straps. Motorized actuators connect to the fingertips, allowing precise monitoring and control of finger movement. Powered by an ESP32 (dual-core) with FreeRTOS, it provides real-time PID-based force feedback at 100Hz. By adjusting PID-setpoints, gripper positions can be mapped to the finger range. A small dither motion reduces stiction (static friction), ensuring smoother force transmission. Communication occurs through a PC linking the ESP32 (via Serial) and robot (via TCP). However, this proof-of-concept is set for future improvements and expanded functionalities.

This work is supported through the InnoFaktur project by the European Regional Development Fund under grant EFRE-20500002. We sincerely thank Professor Dr. Matthias Böhmer for his academic guidance. We also appreciate Cologne UAS and Code & Context for generously supplying materials and resources. Lastly, we thank the Innovation Hub for providing access to their infrastructure.

Keywords: Software Tools for Benchmarking and Reproducibility, Human-Robot Collaboration, Software, Middleware and Programming Environments
Abstract: AURORA is a software platform, that facilitates scalable deployment of robotic simulations over the web for the Human-Robot Interaction (HRI) community. As robotics is becoming increasingly important in various disciplines, there is a growing need for accessible and scalable research methods. Traditional experiments often require expensive hardware and in-person participation, limiting accessibility and participant diversity. Our platform allows researchers from different fields to easily provide HRI experiences by deploying online studies with robotic simulations paired with customizable surveys, allowing end users worldwide to interact with these simulations. Our platform is entirely open source and can be hosted locally, providing flexibility and control of the research environment. Since AURORA is implemented with Docker, it is platform-independent. By offering a user-friendly interface that can be deployed and used without extensive technical expertise, our platform reduces costs, increases participant diversity, and improves the reproducibility of research in the HRI community.