Keywords: AI-Based Methods, AI-Enabled Robotics, Assembly
Abstract: Large Language Models (LLMs) are widely used for long-horizon task planning, often guided by visual demonstrations and online videos. However, visual data alone struggles to capture subtle contact interactions and force-related parameters essential for real-world execution. This paper presents an in-context learning framework that integrates tactile and force-torque information from human demonstrations to enhance LLM-based planning. A bootstrapped reasoning pipeline sequentially incorporates each modality into a task plan, which serves as a reference for new scenarios. Experiments on sequential manipulation tasks validate improved multi-modal understanding and planning performance.

Keywords: Dexterous Manipulation, In-Hand Manipulation, Deep Learning in Grasping and Manipulation
Abstract: This abstract aims to spark a discussion on the key building block for dexterous manipulation: is it learning or compliance? While those are not the only building blocks, both have driven significant progress and merit discussion. An essential factor in addressing this question is evaluating both the generality of a solution and the cost associated with achieving this generality. To compare the two, this abstract looks at one axis of generality: the ability to execute a manipulation skill in different wrist orientations. We show that a compliant hand can perform a cuboid rotation skill in varying wrist orientations at no additional cost. We explain that compliance enables self-stabilization, making it an ideal low-level building block for robust manipulation.

Keywords: Telerobotics and Teleoperation, Industrial Robots, AI-Enabled Robotics
Abstract: Teleoperation using haptic feedback allows users to interact with remote environments while retaining a sense of touch. However, the stability and transparency of these systems are compromised under communication network delay. As a consequence, tasks goals become harder or even impossible to accomplish. This paper explores how the Digital Twin (DT) of a teleoperation system can be used to facilitate the completion of teleoperated industrial assembly tasks. On one hand, we present a vision-based algorithm for rapid 3D object and dynamic environment tracking within Model-Mediated Teleoperation (MMT), enabling users to receive haptic feedback in structured dynamic environments while maintaining robustness against network delays. A user study comparing the proposed method with teleoperation using the Time Domain Passivity Approach (TDPA) was conducted. On the other hand, the DT is used together with human demonstrations to estimate Virtual Fixtures (VFs) that guide the user towards the goal, further simplifying the task at hand. The results demonstrate that MMT exhibits robustness to varying delays in task performance and outperforms TDPA at higher delay levels. The VFs reduce the completion time of the tasks and improve their accuracy.

Keywords: Soft Robot Materials and Design, Soft Sensors and Actuators, Biomimetics
Abstract: The creation of ever more life-like robots has enabled the exploration and monitoring of various ecosystems, greatly contributing to research in ecology and climate change. By harnessing biomimicry, these robots are designed to blend seamlessly into natural environments, utilizing soft materials and artificial muscle technologies. However, sustainability aspects have been widely neglected in the process of developing versatile robots for exploration. To prevent pollution and damage to ecosystems after the end-of-use of a robot, the use of eco-friendly materials in embodiments of robots has become critical. The main challenge in this endeavor lies in merging high performance—such as durability and desirable mechanical and electrical properties—with environmentally friendly design or biodegradability. Here, we focus on soft robots designed for deployment in nature, such as for aquatic environments, and address current and future challenges in utilizing of sustainable materials. In particular, we introduce swimming robots equipped with soft actuators with high performance. Building on this example, we delve into sustainable soft actuators and how they can be integrated in biomimetic robots. We show that with combining materials development and soft actuation designs we can create devices with longer lifetimes that still degrade naturally after use. Performance and sustainability can both be achieved in new technologies if we make the effort to design materials and devices accordingly.

Keywords: Medical Robots and Systems, Sensor Fusion, Computer Vision for Medical Robotics
Abstract: Ultrasound imaging, being an instant, innocuous, and comparatively inexpensive medical imaging modality, is of high value in clinical practice. However, conventional, i.e., humanly performed, medical ultrasound imaging is a procedurally complex task where acquired images depend on a multitude of multi-modal actions and observations. This leads to a marked dependence of the provided diagnostic outcome on the sensorimotor skill, expertise, and experience of the respective user, which motivates efforts to assist, automate, and replicate the functional roles of human operators with robotic systems. We propose ConSensUS, a contextual sensing framework that integrates multiple modalities to enhance robotic ultrasound: (a) visual data from an RGB-D camera, (b) haptic feedback from a force/torque sensor and robotic proprioception, (c) spatial localization via optical tracking, and (d) ultrasound imaging. ConSensUS aims to establish correlations between external morphology, applied forces, and internal anatomy, providing a rich dataset for learning-based approaches. Future work will focus on creating a multimodal sonographic atlas and the identification of multivariate correlation models, e.g., between external morphology captured as RBG-D streams, applied contact wrenches, and internal anatomy visualized by ultrasound imaging, which will be utilized for learning-based models and optimizing scanning strategies.

Keywords: Machine Learning for Robot Control, Optimization and Optimal Control
Abstract: This work presents a closed-loop learning framework for Model Predictive Control (MPC) tailored to autonomous systems, ensuring both performance optimization and safety throughout the learning phase. By parameterizing the MPC stage cost function, the approach enhances adaptability, allowing controllers to balance energy efficiency, trajectory accuracy, and robustness in dynamic environments. A Bayesian optimization-based method iteratively refines the cost function parameters while enforcing probabilistic stability guarantees, leveraging safe learning strategies for efficient exploration of the parameter space. The proposed methodology is demonstrated in robotic applications, including energy-efficient motion planning and control for robotic manipulators and stability-aware learning for highly nonlinear systems, showcasing its effectiveness in achieving provably stable and high-performance control.

Keywords: AI-Enabled Robotics, Wearable Robotics, Deep Learning for Visual Perception
Abstract: In humans vision plays a fundamental role in guiding adaptive locomotion. Similarly, when designing the control strategy for a walking assistive technology, the use of computer vision may substantially improve modulation of the assistance based on the external environment. In this work, we developed a hip exosuit controller able to distinguish among three different walking terrains through the use of an RGB camera and to adapt the assistance accordingly. The system was tested with seven healthy participants walking throughout an overground path comprising of staircases and level ground. Subjects performed the task with the exosuit disabled (Exo Off), constant assistance profile (Vision Off), and with assistance modulation (Vision On). Our results showed that the controller was able to classify in real-time the path in front of the user with an overall accuracy per class above the 85%, and to perform assistance modulation accordingly. Evaluation related to the effects on the user showed that Vision On was able to outperform the other two conditions: we obtained significantly higher metabolic savings than Exo Off, with a peak of ≈ −20% when climbing up the staircase and ≈ −16% in the overall path, and than Vision Off when ascending or descending stairs. Such advancements in the field may yield to a step forward for the exploitation of lightweight walking assistive technologies in real-life scenarios.

Keywords: Swarm Robotics, Human-Robot Teaming, Multi-Robot Systems
Abstract: This paper investigates the ability of multiple operators to dynamically share the control of robot swarms and the effects of different communication types on performance and human factors. A total of 52 participants completed an experiment in which they were randomly paired to work together in guiding the swarm to complete spatially distributed tasks. Results show that although the ability to share robots did not necessarily increase task scores, it allowed the operators to switch between working independently and collaboratively, reduced the total energy consumed by the swarm, and was considered useful by the participants. We validate the sharing of robots among two operators using physical robots, demonstrating its applicability in the real world.

Keywords: Imitation Learning, Surgical Robotics: Laparoscopy, Surgical Robotics: Planning
Abstract: Integrating (semi-)autonomous robotic support into minimally invasive surgery could help address the shortage of surgical personnel and lessen the workload of surgeons. However, these robots must be capable of performing complex tasks within the unpredictable and unstructured settings that characterize surgical environments. In our approach we combine visual features based on surgical domain knowledge with a feedforward neural network for end point prediction and probabilistic movement primitives for trajectory determination. We show that we can effectively learn the retraction task for the removal of the gallbladder based on 200 demonstrations. We evaluated our proposed imitation learning method on both a silicone liver phantom and textit{ex-vivo} porcine livers. With success rates of 91% and 86% for gallbladder retraction, this robotic system effectively supports surgeons during these interventions.

Keywords: Machine Learning for Robot Control, Integrated Planning and Learning, Optimization and Optimal Control
Abstract: Traditional robotic control systems rely on modular sense-plan-act architectures or end-to-end deep learning, both of which struggle with seamless task transfer, real-time adaptation, and computational efficiency. In this work, we propose World Model Predictive Control (WMRC), a novel framework that integrates World Models with Model Predictive Control (MPC) principles. By leveraging pre-trained differentiable world models to predict system dynamics and optimize control actions, WMPRC eliminates the need for extensive policy training, unlike reinforcement learning (RL)-based world models. Our approach unifies multimodal state representations, task-specific cost learning, and constraint-aware optimization within a receding horizon framework. Inspired by human motor control and learning, it integrates general scene understanding and basic dynamics estimation while fine-tuning actions through rapid interaction with the environment. Furthermore, elastic model updating balances short-term corrections—instantaneous reactive adjustments to new dynamics—with long-term knowledge retention, enabling memory-augmented fine-tuning and improved general skill proficiency.

Keywords: Force and Tactile Sensing, Perception-Action Coupling, Reinforcement Learning
Abstract: Humans make extensive use of haptic exploration to map and identify the properties of the objects that we touch. Also, in robotics, the use of active tactile perception has emerged as an important research domain that complements vision for tasks such as object classification, shape reconstruction, and manipulation. In this work, we introduce TAP (Task-agnostic Active Perception) – a novel framework that leverages reinforcement learning (RL) and transformer-based architectures to address the challenges posed by partially observable environments. TAP integrates Soft Actor-Critic (SAC) and CrossQalgorithms within a unified optimization objective, jointly training a perception module and decision-making policy. By design, TAP is task-agnostic and can, in principle, generalize to any active perception problem. We evaluate TAP across diverse tasks, including toy examples and a realistic application involving haptic exploration of 3D models of handwritten digits. Experiments demonstrate the efficacy of TAP, achieving a classification accuracy of 92% on Tactile MNIST. These findings underscore the potential of TAP as a versatile and generalizable framework for advancing active tactile perception in robotics.

Keywords: Grasping, RGB-D Perception, Domestic Robotics
Abstract: We present a fabric grasping pipeline which adapts the type of grasp automatically to the given situation. Thus it is capable of generating grasp poses to grasp folds in the fabric when they are available. Otherwise it will pinch the fabric against a surface to create a fold which is then grasped. The system is demonstrated using an interactive clothes grasping scenario.

Keywords: Grippers and Other End-Effectors, Embedded Systems for Robotic and Automation
Abstract: We present a novel gripper design family which combines low-cost with multimodal sensing and high repairability. This is achieved by relying on 3D-printing and commonly available off-the-shelf components. Replacement parts are either cheap and easily accessible or can by printed on demand. A set of sensor options allows to capture a diverse set of data types during manipulations.

Keywords: Human-Centered Robotics, Surgical Robotics: Planning, Surgical Robotics: Steerable Catheters/Needles
Abstract: The integration of robotic systems into operating rooms (ORs) presents challenges, particularly in ophthalmic surgery, where micrometer-level precision is critical. Procedures like subretinal injections require accurate robotic alignment, yet current methods are time-intensive and inefficient. The ForNeRo project addresses these issues by combining preoperative planning, augmented reality (AR)-based guidance, and advanced robotics. Using Unity3D, a dynamic digital twin of the OR is developed to automate robot placement while accounting for spatial constraints, personnel, and equipment. AR technologies, such as HoloLens 2, enable real-time visualization and precise robot positioning during surgery. Collaborative efforts by FAU, TUM, and Custom Surgical focus on addressing ophthalmology’s unique demands. Preliminary results indicate reduced setup times and improved surgical accuracy. Future work includes expanding digital twin capabilities, integrating predictive AI algorithms, and conducting clinical validation to advance robotic integration in surgery

Keywords: Biologically-Inspired Robots, Biomimetics, Compliant Joints and Mechanisms
Abstract: This paper introduces the bio-inspired open joint for dynamic robots. The joint structure features the integration of the open joint and ligamentous constraint. The proposed structure is made from relatively simple machined components and can be easily repaired. Verification of the proposed approach includes the qualitative evaluation of ROM (Range of Motion). The tests showed that the proposed joint structure can be a relevant approach for replicating the ROM of a human knee joint. Future works include the actuation of the joint system and integration into a real robot.

Keywords: Wearable Robotics, Rehabilitation Robotics, Human Performance Augmentation
Abstract: Human locomotion is a highly adaptive and efficient process, influenced by biomechanical, neuromuscular, and strength-related factors. However, aging and neuromuscular impairments can negatively impact walking efficiency and transition speeds. This study investigates the Biarticular Thigh Exosuit (BATEX), a soft wearable device designed to assist both hip and knee joints through targeted actuation. By supporting human rectus femoris and hamstring muscles, BATEX aims to improve Preferred Walking Speed (PWS) and Preferred Transition Speed (PTS). An experimental study with 12 participants evaluated the impact of BATEX on gait dynamics. Results showed a 14% increase in PWS and a 9% increase in PTS, demonstrating that BATEX effectively enhances locomotor efficiency while reducing neuromuscular effort. These findings suggest potential applications in human-robot interaction in healthcare, e.g., mobility assistance, training, and rehabilitation robotics

Keywords: Aerial Systems: Mechanics and Control, Aerial Systems: Applications, Aerial Systems: Perception and Autonomy
Abstract: In the evolving landscape of high-speed agile quadrotor flight, achieving precise trajectory tracking at the platform's operational limits is paramount. Controllers must cope with highly nonlinear dynamics and actuator constraints while being robust to disturbances and uncertainties for real-world applications. In this work, we tightly couple sensor-based control with optimal model predictive control to achieve robust agile flight. The controller addresses individual limitations of existing state-of-the-art control paradigms and unifies their strengths. We highlight the controller’s potential for accurate tracking of highly aggressive trajectories that surpass the feasibility of the actuators in exceedingly disturbance-prone settings. The combination of incremental control and convex optimization yields a robust and highly-efficient controller. Notably, the online execution avoids matrix factorizations and divisions enabling fast computation at high update rates and numerically stable execution.

Keywords: Imitation Learning, Incremental Learning, Learning from Demonstration
Abstract: Robots should be able to continuously learn and enhance their skills and abilities over time. Such skills can be represented as sequences of movement primitives, which are known for their generalization abilities and can be re-used across tasks when being collected in a library. Incrementally learning such a library efficiently allows movement primitives to be updated without the need to permanently store all demonstrations. This extended abstract builds on our previous work on incrementally learning Full-Pose Via-Point Movement Primitives based on 7 fundamental operations. Here, we integrate this movement primitive library into a cognitive architecture and conceptionally show how to incrementally learn such a library for a humanoid robot from demonstrations of multiple modalities.

Keywords: Imitation Learning, Modeling and Simulating Humans, Bioinspired Robot Learning
Abstract: Human walking is a highly coordinated process that requires precise muscle activation to ensure stability and efficiency. Simulating realistic gait patterns is crucial for applications such as gait assistance with exoskeletons, prosthetics, and rehabilitation technologies. This study presents an imitation learning-based framework that utilizes the GAIL algorithm to generate human-like walking by optimizing metabolic cost. The model, trained using kinematic data, produces muscle activation patterns comparable to real EMG patterns and OpenSim’s Computed Muscle Control tool. The key findings show correlations of 0.95 for Gastrocnemius, 0.89 for Soleus, and 0.79 for Vastus, indicating a strong agreement between the learned and simulated activations. While certain muscles, such as the Biceps Femoris, exhibit higher discrepancies, the overall results demonstrate that imitation learning can achieve physiologically plausible gait patterns without relying on external force data. Future work will include comparisons with experimental EMG recordings and refinements to the reward function by incorporating distributed forces and co-contraction handling.

Keywords: Optimization and Optimal Control, Embedded Systems for Robotic and Automation, Machine Learning for Robot Control
Abstract: Model Predictive Control (MPC) is a powerful framework for robotic motion control, offering optimal performance and constraint satisfaction. However, its high computational demands often hinder real-time deployment on embedded systems. In this work, we propose a learning-based approach to approximate MPC laws using Gaussian processes (GPs), significantly reducing online computational complexity while retaining control performance. By training GPs on optimally generated MPC trajectories, we achieve a 100× speedup in computation time compared to solving MPC optimization problems online. We demonstrate the effectiveness of our approach on a quadrotor flight control task, showing that the GP-based controller closely mimics the behavior of the full MPC while being orders of magnitude more computationally efficient. Our results highlight the potential of Gaussian process approximations for enabling real-time, high-performance control on resource-constrained robotic platforms.

Keywords: Telerobotics and Teleoperation, Physical Human-Robot Interaction, Human-Robot Collaboration
Abstract: Teleoperation is a crucial technology in the modern era, complementing robotic systems while serving as a vital fallback solution in the event of a critical autonomous system failure. This technology is instrumental in various fields, including robotic-assisted minimally invasive surgery, hazardous environment handling, telenavigation, space exploration, military operations, underwater exploration, industry, and search and rescue operations. A critical challenge in teleoperation is the presence of communication time-delays, which can significantly impact system stability, control performance, operator safety, and overall efficiency. These delays arise due to the inherent latency in transmitting control commands and receiving force feedback over communication networks. Even minor time-delays can lead to severe destabilization effects. Additionally, issues such as limited force feedback, bandwidth constraints, human operator fatigue, and restricted situational awareness further complicate teleoperation implementations. To address these challenges, extensive research has been conducted at German Aerospace Center (DLR) to develop various robust control strategies to enhance the transparency, robustness, and safety of teleoperation systems across different applications and robotic platforms.

Keywords: Collision Avoidance, Robust/Adaptive Control, Optimization and Optimal Control
Abstract: Collision avoidance in autonomous vehicles is particularly challenging in uncertain, dynamic environments. This work presents a robust learning-based collision avoidance framework that integrates data-driven reachability analysis with Model Predictive Control (MPC) to handle uncertain moving obstacles. Designed for scenarios with limited obstacle information, the method learns over-approximations of obstacle reachability using zonotopes, incorporating a safety margin learned from past uncertain measurements with bounded noise. The learned reachability sets are embedded as polytopic collision avoidance constraints within a robust MPC framework, ensuring reliable vehicle control while guaranteeing impact-free navigation. Simulation results confirm the effectiveness of our approach, demonstrating safe and adaptive learning-based collision avoidance for autonomous vehicles in uncertain environments.

Keywords: Robust/Adaptive Control, Autonomous Vehicle Navigation, Autonomous Agents
Abstract: Machine learning techniques, particularly neural networks, offer significant potential to enhance the performance and applicability of Model Predictive Control (MPC) in real-world systems. However, their integration poses challenges due to their inherent unpredictability, lack of formal performance guarantees, and susceptibility to errors. Conventional robust MPC approaches mitigate these risks by introducing safety margins around neural network-supported predictions, but this often leads to overly conservative control actions and, in some cases, infeasibility. To address this, we propose a safe and robust neural network-assisted MPC framework that enforces bounded network outputs, ensuring that predictions remain close to a nominal model. By directly incorporating error dynamics into the network’s output function, our approach eliminates the need for additional control inputs to compensate for uncertainties, thereby avoiding unnecessary input constraint tightening. The method guarantees constraint satisfaction, ensures robust set stability, and provides explicit worst-case performance bounds in the event of neural network malfunctions. We validate the proposed approach through an autonomous rover operating in an uncertain environment, demonstrating its ability to maintain safety, stability, and improved performance while reducing conservatism compared to traditional robust MPC methods.

Keywords: Humanoid and Bipedal Locomotion, Reinforcement Learning
Abstract: In this work, we present an approach to reliably learn stable gait patterns for a self-designed biped robot with Reinforcement Learning (RL) in the simulation environment Mujoco. For that we designed a model of the robot, which captures its kinematic structure, including elastic components and closed kinematic structures. The reward function, as well as the agents observation, is based on a clock signal which is a key component to learn symmetrical, stable and humanlike gait patterns with multiple RL algorithms reliably and fast. To make the biped robot turn, we follow a multi-task-phase approach, where we separate between phases of straight walking, turning and transition phases.

Keywords: AI-Enabled Robotics, Data Sets for Robotic Vision, Data Sets for Robot Learning
Abstract: This work presents a novel pipeline for generating diverse image datasets using inpainting techniques to address data scarcity in Visual Question Answering (VQA) tasks. By fine-tuning TinyLLaVA with Low-Rank Adaptation on 6,000 generated images, we achieve improved performance on VQA v2 compared to the base model. FID scores validate our generated data’s quality, suggesting potential applications for enhancing datasets across various domains where data collection is challenging. Our approach offers a scalable solution for dataset augmentation in multimodal learning.

Keywords: Telerobotics and Teleoperation, Safety in HRI, Compliance and Impedance Control
Abstract: Safety is a critical aspect whenever robots interact with humans. This also applies to the field of telemanipulation, where the operator might not be fully aware of the remote surroundings. Thus, improvements are necessary to enable telemanipulation in various application areas. In this work, we present our previously proposed algorithms to improve safety and usability in telemanipulated human-robot interaction tasks. It includes an interaction force dependent variable impedance controller to build a safety mechanism and a full 6D variable virtual fixture algorithm to increase accuracy in different applications. The results show great improvements in task performance, system safety and mental workload of the operator.

Keywords: Human-Centered Robotics, Modeling, Control, and Learning for Soft Robots, Motion Control
Abstract: The human wrist exhibits multi-degree-of-freedom motion and adaptive impedance that are challenging to replicate in prosthetic devices. This paper presents a novel flexible wrist design based on dual Handed Shearing Auxetic (HSA) cylinders. The design provides two essential degrees of freedom—flexion/extension (F/E) and pronation/supination (P/S)—with a near-linear impedance response. Finite element analysis indicates that servo actuation within ±45◦ optimally drives the mechanism. Experiments on a dedicated test platform reveal strong linear correlations (R2 > 0.99) between servo input and wrist output (approximately 25◦ for F/E and 29◦ for P/S) and driving torques exceeding 3.7 N·m, validating the design’s potential for enhanced prosthetic wrist functionality.

Keywords: Human-Robot Collaboration, Wearable Robotics
Abstract: Active exosuits and exoskeletons require advanced control strategies to ensure seamless human integration, safety, and adaptability. Effective interaction between the biological body and assistive robots demands smooth, cooperative, and constraint-aware control mechanisms. This paper presents the development and evaluation of a Model Predictive Control (MPC) system for the BATEX exosuit, focusing on improving the control of Series Elastic Actuators in assistive devices. We compare MPC with traditional control strategies and demonstrate its superior ability to handle dynamic conditions, ensure constraint satisfaction, and enhance user comfort. Results indicate that MPC provides smoother force control and improved disturbance handling while maintaining safe and adaptive human-exoskeleton interaction.

Keywords: Aerial Systems: Perception and Autonomy, Aerial Systems: Applications, Field Robots
Abstract: This paper presents the key outcomes of project WildCap (2021–2026), funded by Cyber Valley, Germany. WildCap's goal is to develop robotic systems and methods to monitor endangered wild animals without physical interference. In this project, we have designed (i) autonomous, vision-based aerial robots that detect, track, and follow animals, and (ii) learning-based methods for animal behavior inference and pose and shape estimation. Our robots, including helium-based airships and multi-rotor drones, use autonomous decision-making to optimize flight paths for accurate animal tracking. The WildCap system was successfully tested at the Pentezug reserve, Hortobágy National Park, Hungary, where it autonomously tracked endangered Przewalski’s horses in a 3,000-ha semi-wild environment. Our behavior inference method was demonstrated on Grévy’s zebras at Mpala Conservancy, Kenya, using drone-acquired videos to estimate long-term activity budgets (e.g., grazing vs. walking) and movement patterns. Such insights help local authorities manage human activities, ensuring designated grazing areas while minimizing wildlife conflict. Additionally, we conducted flight experiments to assess drone noise impacts on equids and other species in the same landscape. This research advances non-invasive wildlife monitoring at scale, aiding conservation and sustainable land management.

Keywords: Prosthetics and Exoskeletons, Grasping, Soft Robot Applications
Abstract: Hand prostheses face the inherent trade-off between dexterity, which is enabled by many degrees of actuation, and intuitive use, which demands both system robustness and simplicity of control. In this work, we present a soft hand prosthesis with embodied multi-grasp synergistic activation. We introduce a novel valve, which allows to switch between precision grasp, tripod grasp and power grasp postures based on the required grasping force. The selection of the valve’s design parameters enables the customization of synergistic activation forces and facilitates the control of multiple grasping patterns by modulating the intensity of a single neural input.

Keywords: Mobile Manipulation, Whole-Body Motion Planning and Control, Motion Control
Abstract: Mobile manipulation tasks usually entail a sequence of sub-tasks to be taken in the given order. Typical approaches execute each action separately, resulting in the robots coming to a complete stop. In this work, we propose a novel Hierarchical-Task Model Predictive Control framework that leverages the robots' redundancy and improves their performance and reactivity for sequential tasks. We showed through real-robot experiments that our proposed hierarchical task control architecture enables the robot to traverse a shorter path in task space and achieves an execution time 2.3 times faster than the typical single-task approach.

Keywords: Motion and Path Planning, Human-Robot Collaboration, Human Detection and Tracking
Abstract: This study presents a MPC-based human-following control system for a differential-drive mobile robot, designed to enable precise path recording and replication in teach-and-repeat (T&R) programming. The system enables a robot to accurately follow a human’s path by recording and replicating their movements. The proposed solution addresses the limitations of existing human-following systems, which suffers from "short-cutting" behavior in curves, which make person following challenging in narrow environments. The system consists of two main components: a hybrid tracking system using UWB and vision-based positioning fusion and a following control algorithm. The following control algorithm is based on Model Predictive Control (MPC), which optimizes the robot's velocity commands to minimize deviations from the human's path while considering kinematic constraints. The MPC predicts the robot's future states over a prediction horizon and computes optimal control inputs to guide the robot along the reconstructed path. The path reconstruction process involves recording the human's trajectory, applying a smoothing algorithm and updating the path in real-time. The proposed system was evaluated through experimentals using a two wheeled balancing robot, demonstrating superior accuracy of path repeating compared to previous approaches. Results indicate significant improvements in path deviation reduction and short-cutting mitigation, confirming the system's effectiveness in T&R applications.

Keywords: Swarm Robotics, Energy and Environment-Aware Automation, Embedded Systems for Robotic and Automation
Abstract: The utility of swarms of robots would greatly increase if they could operate over extended periods of time. Here, we consider two strategies for swarms of robots to replenish their energy while performing work in a remote location. In the first, each robot commutes to work and replenishes at its base. In the second, some robots perform work, whereas others commute to provide them with energy. We present results from extensive physics-based simulations. The first strategy performs 92.8% of the work at only 12.6% lower energy efficiency than an optimal strategy. The second strategy is beneficial for low charging rates or if the robots providing energy are permitted increased amounts of storage. We provide proof-of-concept validation using the CapBot swarm robot platform.

Keywords: Sensor Fusion, Sensorimotor Learning, Reinforcement Learning
Abstract: Humans seamlessly integrate multiple sensory in- puts to perform complex manipulation tasks, adapting to sensory perturbations and dynamic changes. In contrast, robotic reinforcement learning (RL) struggles with pretraining from heterogeneous sensor modalities due to varying sensory feature distributions and their changing importance depending on the task phase. We propose MultiSensory Dynamic Pretraining (MSDP), a novel framework for learning multisensory representations tailored for task-oriented policy learning using masked autoencoding and self-supervised objectives to shape and fuse sensory features. Evaluation on a challenging 2d box pushing and a contact-rich manipulation task showcase the effectiveness of MSDP. Our framework’s modular pretraining process supports various sensor combinations, providing a simple and effective solution for complex manipulation tasks.

Keywords: Calibration and Identification
Abstract: Based on our recent work [4], an extension to the state-of-the-art nonlinear least squares optimization approach for multi-sensor hand-eye calibration [1] is provided. We consider a constraint on the relative poses between coordinate frame triplets by implicitly enforcing a closed-loop pose-graph without introducing additional unknowns like proposed in [5] and [6]. We show preliminary experimental results on synthetic data for the minimal setting of two sensors attached to a robot arm forming one minimal closed loop between three frames.

Keywords: Learning from Demonstration, Deep Learning Methods, Deep Learning in Grasping and Manipulation
Abstract: Spatial understanding is a critical aspect of most robotic tasks, particularly when generalization is important. Despite the impressive results of deep generative models in complex manipulation tasks, the absence of a representation that encodes intricate spatial relationships between observations and actions often limits spatial generalization and sample efficiency. To tackle this problem, we introduce a novel policy class, ActionFlow. ActionFlow integrates spatial symmetry inductive biases while generating expressive action sequences. ActionFlow introduces an SE(3) Invariant Transformer architecture for spatial reasoning in SE(3). For action generation, ActionFlow leverages Flow Matching to generate high-quality samples with fast inference. In combination, ActionFlow policies exhibit strong spatial and locality biases and SE(3)-equivariant action generation. Our experiments demonstrate the effectiveness of ActionFlow and its two main components on simulated and real-world robotic manipulation tasks and confirm that ActionFlow yields equivariant, accurate, and efficient policies.

Keywords: Sensor Fusion, Sensorimotor Learning, Representation Learning
Abstract: Humans develop a body representation based on proprioceptive, tactile and visual information allowing them to adapt to bodily changes without external supervision or access to a prespecified body plan. We propose a biologically inspired learning framework that enables robots to autonomously develop a body schema using only sensory observations. After learning, our robot shows human-like sensory integration patterns in the Rubber Hand Illusion experiment highlighting its ability to recalibrate its body schema by flexibly integrating conflicting sensory signals.

Keywords: Machine Learning for Robot Control, Learning from Demonstration, Motion and Path Planning
Abstract: Ensuring safe and efficient motion planning in dynamic and uncertain environments remains a fundamental challenge in robotics. Diffusion-based planners have emerged as powerful generative models for encoding and synthesizing feasible paths from demonstrations, yet they lack guarantees on constraint satisfaction and real-time adaptability. In this work, we explore the integration of diffusion-based path planning with model predictive control to ensure safe and adaptive motion execution. We first review existing work on both approaches and then propose their combination, leveraging path-following MPC to adaptively track diffusion-generated paths while ensuring constraint satisfaction. This synergy enables robust, data-driven motion planning with real-time feasibility and safety guarantees.

Keywords: Human Detection and Tracking, Sensor Fusion, Localization
Abstract: How teams collaborate to perform complex tasks, from team sports to surgical procedures, has previously been investigated via multimodal sensing and analysis. Ultra-wideband (UWB) positioning systems are highly mobile and can be used to track collaborating team members even in cramped environments. However, the sampling rate of UWB systems is inversely proportional to the number of people tracked, and their accuracy is hindered by electromagnetic occlusion. To improve position and orientation estimation during team collaborative studies, we propose to fuse UWB positioning with a wearable inertial measurement unit (IMU) %measurements by applying an error-state extended Kalman filter (ES-EKF). This filter offers faster and more consistent estimation and remains functional even in the absence of UWB input. Single-human and multi-human sessions were recorded and filtered for evaluation against ground truth from optical motion capture. By integrating IMU readings, the ES-EKF increases the sampling rate from 0.5–20 Hz to 100 Hz. Even by correcting only planar position in the room, the ES-EKF yields improved results over UWB in four out of six DOF: lateral and longitudinal position and yaw and pitch orientation.

Keywords: Integrated Planning and Control, Motion Control, Optimization and Optimal Control
Abstract: Tuning controller parameters in autonomous systems is time-consuming and costly, requiring extensive real-world testing and high-fidelity simulations. To address this, we propose a multi-fidelity Bayesian optimization framework that integrates simulation-based tuning with experimental validation, leveraging a linear auto-regressive Gaussian process. By exploiting correlations between low-cost simulations and high-fidelity experiments, our approach improves sample efficiency without requiring simultaneous evaluations at all fidelity levels. We validate our method on an autonomous vehicle tracking controller, demonstrating reduced computational and experimental costs while ensuring high-performance parameter selection.

Keywords: Optimization and Optimal Control, Probability and Statistical Methods, Reinforcement Learning
Abstract: Efficient global optimization and sampling remain fundamental challenges, particularly in areas such as robotics and reinforcement learning, where gradients may not be available or reliable. While Stein Variational Gradient Descent (SVGD) provides a powerful framework for sampling diverse solutions, its reliance on gradient information limits its applicability. Existing gradient-free SVGD variants often suffer from slow convergence, and poor scalability. To improve gradient-free sampling and optimization, we propose Stein Variational CMA-ES, a novel gradient-free inference method that combines the efficiency of evolution strategies with SVGD-based repulsion forces. Empirical evaluation across density estimation and continuous control tasks demonstrates that our approach efficiently adapts search distributions while maintaining solution diversity. Our findings establish SV-CMA-ES as a scalable method for zero-order density approximation and blackbox optimization, bridging the gap between SVGD and evolution strategies.

Keywords: Machine Learning for Robot Control, Reinforcement Learning
Abstract: Manipulating objects with varying geometries and deformable objects is a major challenge in robotics. Tasks such as insertion with different objects or cloth hanging require precise control and effective modelling of complex dynamics. In this work, we frame this problem through the lens of a heterogeneous graph that comprises smaller sub-graphs, such as actuators and objects, accompanied by different edge types describing their interactions. This graph representation serves as a unified structure for both rigid and deformable objects tasks, and can be extended further to tasks comprising multiple actuators. To evaluate this setup, we present a novel and challenging reinforcement learning benchmark, including rigid insertion of diverse objects, as well as rope and cloth manipulation with multiple end-effectors. These tasks present a large search space, as both the initial and target configurations are uniformly sampled in 3D space. To address this issue, we propose a novel graph-based policy model, dubbed Heterogeneous Equivariant Policy (HEPi), utilizing SE(3) equivariant message passing networks as the main backbone to exploit the geometric symmetry. In addition, by modeling explicit heterogeneity, HEPi can outperform Transformer-based and non-heterogeneous equivariant policies in terms of average returns, sample efficiency, and generalization to unseen objects.

Keywords: Human-Robot Collaboration, Motion and Path Planning, Grasping
Abstract: Human-Robot Collaboration (HRC) scenarios demand computationally efficient frameworks that enable natural and safe actions and interactions in shared workspaces. To address this, we propose a novel framework that utilises interactive Gaussian Process (GP) distance fields applying Riemannian Motion Policies (RMP) for key HRC functionality. We demonstrate the versatility of our CPU-only system in common HRC scenarios where a collaborative robot (cobot) interacts safely and naturally with a human and performs grasping actions in a dynamic environment. By providing a continuous and differentiable distance field and combining motion generation, obstacle avoidance, and object manipulation within a single system, we aim to broaden the scope and accessibility of HRC research in real dynamic environments.

Keywords: Machine Learning for Robot Control
Abstract: Maximum entropy reinforcement learning (MaxEnt-RL) has become the standard approach to RL due to its beneficial exploration properties. Traditionally, policies are parameterized using Gaussian distributions, which significantly limits their representational capacity. Diffusion-based policies offer a more expressive alternative, yet integrating them into MaxEnt-RL poses challenges—primarily due to the intractability of computing their marginal entropy. To overcome this, we propose Diffusion-Based Maximum Entropy RL (DIME). DIME leverages recent advances in approximate inference with diffusion models to derive a lower bound on the maximum entropy objective. Our method enables the use of expressive diffusion-based policies while retaining the principled exploration benefits of MaxEnt-RL, significantly outperforming other diffusion-based methods on challenging high-dimensional control benchmarks. It is also competitive with state-of-the-art non-diffusion based RL methods while requiring fewer algorithmic design choices and smaller update-to-data ratios, reducing computational complexity.

Keywords: SLAM, Mapping, Data Sets for Robotic Vision
Abstract: This paper addresses the problem of rotational motion estimation using event cameras. Several event-based rotation estimation methods have been developed in the past decade, without considering a global refinement step. To this end, we propose the first event-based rotation-only bundle adjustment (BA) approach. We formulate it leveraging the state-of-the-art Contrast Maximization (CMax) framework, which is principled and avoids the need to convert events into frames. In addition, we use the proposed BA to build CMax-SLAM, the first event-based rotation-only SLAM system comprising a front-end and a back-end. Our BA is able to run both offline (trajectory smoothing) and online (CMax-SLAM back-end). We demonstrate the proposed method in a variety of scenarios, which shows state-of-the-art performance. Project page and full paper: https://github.com/tub-rip/cmax_slam

Keywords: Software Architecture for Robotic and Automation, Distributed Robot Systems, Software-Hardware Integration for Robot Systems
Abstract: Deploying AI-powered autonomous robots for 24/7 operation is still very challenging due to the resource demands, the high complexity of the integrated system, and the reliability and robustness requirements. In this paper, we briefly present our research activities on cloud native robotics with a focus on robotic software development and deployment from a system perspective. Considering the requirements of reliability, adaptability, and scalability, we have proposed approaches in previous works to take advantage of cloud native computing to help the AI-based robotic system to work for the whole day operation. Using core tasks such as navigation, motion planning, and manipulation, we have demonstrated the benefits in terms of scalability and usability. Cloud robotics is a cross-cutting topic that is closely related to the application and can primarily improve the efficiency of the entire team. Therefore, we also point out some open challenges from our perspective that need to be addressed through community collaboration.

Keywords: Legged Robots, Biologically-Inspired Robots
Abstract: Miniature robots provide unprecedented access to confined environments and show promising potential for novel applications such as search and rescue and high value asset inspection. In nature, animals demonstrate high levels of robustness to obstructions and damage. In biology animals use various different mechanisms to enable agile and adaptable locomotion in complex natural terrains. When animals are challenged with confined spaces, the interaction between the environment and body has a significant impact on the ability to navigate efficiently and successfully. Animals use a combination of active neural control and passive environment interaction. Seen across nature, varying body morphologies have a variety of appendage to body ratios. Current robots (especially on the insect scale) cannot modify their shape to significantly improve performance or add new functionality. The capability of body deformation further enhances the reachability of these small robots in complex cluttered terrains similar to those of insects and soft arthropods. Most large scale legged robotics systems are designed with a fixed rigid central body structure containing the computational hardware and most of the actuation. Current robots (especially on the insect scale) cannot modify their shape to significantly improve performance or add new functionality. The capability of body deformation further enhances the reachability of these small robots in complex cluttered terrains similar to those of insects and soft arthropods. Biological inspiration from cockroaches and spiders has led to the development of millimeter-scale legged robots with soft-compliant bodies, to allow novel abilities to walk through previously inaccessible environments.

Keywords: Machine Learning for Robot Control, Model Learning for Control
Abstract: Mobile robots should be capable of planning cost-efficient paths, even if terrain properties, such as friction, vary with location. TRADYN is a terrain- and robot-aware forward dynamics model for a simulated 2D unicycle robot, which is conditioned on a context variable and terrain feature lookups, and can therefore adapt to various environment properties. This extended abstract evaluates the prediction and planning performance of TRADYN under different sources of noise. We find that terrain lookup and especially calibration have a positive effect on prediction performance even under noisy conditions.

Keywords: Multi-Robot Systems, Data Sets for Robot Learning, AI-Enabled Robotics
Abstract: A simulation framework is introduced based on ROS and NVIDIA's Isaac Sim to create a scalable, adaptable Industry 4.0 virtual laboratory for the integration and testing of 3D Digital Twin production assets and interactions with multi-robot systems in complex environments. With features like scene generation, virtual sensor definition and interoperable data collection, the pipeline enables research in algorithm design, AI, and human-robot collaboration.

Keywords: AI-Enabled Robotics, Cognitive Control Architectures
Abstract: Recent advancements in AI-powered and cognition-enabled robotics are reshaping the capabilities of robotic systems, enabling them to operate autonomously in dynamic and unstructured environments. This paper explores how neural and probabilistic models, symbolic reasoning frameworks, and adaptive control mechanisms contribute to these advancements. These innovations address scalability, transparency, and reliability while facilitating human-robot collaboration. We examine their transformative potential in key domains such as healthcare, logistics, and manufacturing, and outline future directions to bridge abstract cognitive reasoning with physical execution, setting the stage for a new era of robotic capabilities.

Keywords: Calibration and Identification, Computer Vision for Automation, Field Robots
Abstract: We present a universal calibration toolbox for complex multi-sensor setups of large-scale mobile off-road robots. This user-friendly toolbox works with both ROS 1 and ROS 2 and will be open-sourced for wider accessibility and feature extension.

Keywords: Visual Servoing, AI-Enabled Robotics, Visual Tracking
Abstract: Visual servoing plays a crucial role in robotic autonomy, enabling real-time control based on image feedback, especially in dynamic environments. The interaction of kinematic arms with dynamic environments, particularly in Human-Robot-Interaction scenarios, remains a challenging issue. Traditional methods rely on geometric model recognition, but recent advances in artificial intelligence (AI) offer feature-based approaches to enhance adaptability and robustness. This extended abstract reviews recent visual servoing techniques, focusing on real-time performance and computational challenges. Furthermore, practical insights from AI-integrated servoing systems are discussed.

Keywords: Human-Robot Collaboration, AI-Enabled Robotics, Computer Vision for Automation
Abstract: Modern industrial production faces numerous challenges, increasing the demand for flexible automation solutions. To address these challenges, a wide array of automation approaches and methods have been developed. Among these, human-robot collaboration and programming by demonstration are particularly promising for ensuring flexibility. A critical factor for the widespread adoption of these methods is accurate 6D pose estimation, which has achieved significant advancements recently, largely driven by breakthroughs in AI technologies. However, many challenges remain unresolved. This paper discusses these challenges and provides perspectives on future research needed to overcome these obstacles, ultimately working toward integrating these technologies into a holistic approach for flexible automation through human-robot collaboration.

Keywords: Semantic Scene Understanding, Task and Motion Planning, Learning from Demonstration
Abstract: Understanding how humans execute and sequence their actions is essential for learning robot skills from demonstrations of long-horizon tasks. An efficient approach is to decompose such tasks into smaller sub-tasks composed of atomic skills and some objects, and properly sequenced to complete the overall task. We present a novel Graph-based neural architecture for learning task graphs from human demonstrations. In particular, we train our network to predict the next actions, object-action saliency, and the subsequent evolution of the scene via motion forecasting, i.e., capturing high-level task dynamics. We present some initial results showing the efficacy of our method on various bimanual tasks.

Keywords: AI-Enabled Robotics, Task Planning, AI-Based Methods
Abstract: Robots need to understand their environment and plan actions to complete tasks effectively. While recent approaches combine Large Language Models (LLMs) with traditional planning algorithms to improve reasoning capabilities, they face several limitations: they can't easily dynamically adapt to changes in the environment, may generate unreliable plans due to LLM hallucinations, and are constrained by the closed-world assumption of classical planners. We propose AutoGPT+P, which combines an affordance-based scene representation with a planning system. By deriving planning domains based on affordances - action possibilities of objects offered to an agent in the environment - AutoGPT+P enables symbolic planning with arbitrary objects. Given a task description specified by the user in natural language, it generates and executes plans that handle incomplete information by exploring the scene or suggesting alternatives. AutoGPT+P achieves a 98% success rate on the SayCan instruction set and 79% on a new dataset of 150 complex scenarios, including tasks with missing objects. The code and dataset are available at https://git.h2t.iar.kit.edu/sw/autogpt-p.

Keywords: Deep Learning for Visual Perception, AI-Enabled Robotics, Motion and Path Planning
Abstract: Intelligent robots must understand their surroundings to navigate effectively. This work explores the use of semantic information to enhance sensor data, enabling robots to recognize both known and novel objects through open-world semantic segmentation. Beyond classification, robust navigation requires terrain assessment, including traversability and object dynamics, which are crucial for safe path planning. Since different robot embodiment have unique mobility constraints, encoding embodiment-specific information allows for more adaptive navigation. Additionally, tracking dynamic entities improves interaction and collision avoidance. This research investigates a multimodal approach that combines semantic and embodiment data with online self-supervision in a world model to advance intelligent navigation across diverse environments.

Keywords: Prosthetics and Exoskeletons, Rehabilitation Robotics
Abstract: Many hand afflictions can lead to a reduced range of motion. Treating these conditions requires the supervision of trained physiotherapists. Germany and the wider world face a critical healthcare personnel shortage. A suitable exoskeleton may allow therapists to supervise more patients simultaneously. We are developing a prototype for hand rehabilitation that actively assists all of the thumb's degrees of freedom (DoF), leveraging insights from previous work. The current state of the prototype is presented, and an outlook is given on how this research project will proceed.

Keywords: AI-Based Methods, Data Sets for Robotic Vision, Vision-Based Navigation
Abstract: Terrain traversability estimation is crucial for autonomous robots, especially in unstructured environments where visual cues and reasoning play a key role. While vision-language models (VLMs) offer potential for zero-shot estimation, the problem of traversability classification remains inherently ill-posed. To explore this, we introduce a small dataset of human-annotated water traversability ratings, revealing that while estimations are subjective, human raters still show some consensus. Additionally, we propose a simple pipeline that integrates VLMs for zero-shot traversability estimation. Our experiments reveal mixed results, suggesting that current foundation models are not yet suitable for practical deployment but provide valuable insights for further research.

Keywords: Reinforcement Learning, Motion and Path Planning
Abstract: We propose MoRe-ERL, an episodic reinforcement learning (ERL) framework that learns residual and refines preplanned reference trajectories into safe, feasible, and efficient task-specific trajectories. MoRe-ERL is a general framework that can seamlessly plug in to arbitrary ERL methods and motion generators. It identifies trajectory segments requiring modification while preserving critical task-related maneuvers and then generates smooth residual adjustments using B-Spline-based movement primitives to ensure adaptability to dynamic task contexts and smoothness in trajectory refinement. Experimental results demonstrate that residual learning significantly outperforms training from scratch using ERL methods, achieving superior sample efficiency and task performance. Hardware evaluations further validate the framework, showing that policies trained in simulation can be directly deployed in real-world systems, exhibiting a minimal sim-to-real gap.

Keywords: Haptics and Haptic Interfaces, Telerobotics and Teleoperation, Wearable Robotics
Abstract: Haptic feedback offers innovative ways to provide information to the operator during telemanipulation tasks. In this work, a low-cost, easy to use tactile bracelet that can be included in existing telemanipulation systems is presented. It consists of up to 8 linear resonant actuators that can be controlled using a ROS2 interface. A user-study with 9 participants was performed to investigate its usability. A position detection accuracy of 88.52% shows promising results for implementation into a telemanipulation system. Further, a lower limit for the time discrimination during pulse patterns is defined with 0.18 s.

Keywords: Humanoid and Bipedal Locomotion, Reinforcement Learning, Motion and Path Planning
Abstract: Navigation is a crucial task for humanoid robots. Bipedal robots are inherently unstable and require constant readjustment to remain upright. This paper outlines an approach to planning footstep location using a policy trained with reinforcement learning, executed using a walking engine based on Cartesian splines. We investigate the effect of using different input modalities for the policy. We found that including measurements obtained from an inertial measurement unit (IMU) increased training performance but reduced execution performance. The attached video shows our results.

Keywords: Robot Safety, Machine Learning for Robot Control
Abstract: Control barrier function (CBF)-based safety filters are used to certify and modify potentially unsafe control inputs to a system, such as those provided by a reinforcement learning agent or a non-expert user. Originally designed for continuous-time systems, CBF safety filters typically assume that the system's relative degree is well-defined and is constant across the domain; however, this assumption is restrictive and rarely verified---even linear system dynamics with a quadratic CBF candidate may not satisfy this assumption. When this assumption is not met, the resulting safety filter optimization problem can lead to unconstrained control inputs over a finite time interval, causing chattering issues and constraint violations. In this work, we illustrate this fundamental issue with a simple example and propose an alternative formulation to address these challenges. Our method leverages multiple CBFs to ensure well-posed safety constraints at all times, effectively preventing undesired oscillations and safety violations. The effectiveness of our proposed method is demonstrated through real-world quadrotor experiments.

Keywords: Human-Centered Robotics, Virtual Reality and Interfaces, Human-Robot Collaboration
Abstract: This research proposal investigates how Mixed Reality can enhance intuitive programming of industrial robots to improve accessibility for non-experts. By evaluating different interaction paradigms and display modalities, we aim to optimize usability, flexibility and efficiency in robot programming. In a comparative study, MR-based programming will be evaluated against the conventional programming by demonstration method, focusing on accuracy, adaptability, and overall effectiveness. To validate our approach, we will integrate MR-based programming interfaces and external tracking systems into a robotic welding setup. This system aims to enable accurate path generation and provides empirical data on the feasibility of MR for industrial automation. Our final goal is to develop a user-centered and flexible programming method that minimizes effort while maintaining high accuracy, enabling broader adoption of MR in manufacturing.

Keywords: Compliance and Impedance Control, Safety in HRI, Telerobotics and Teleoperation
Abstract: The Elasto-Plastic Robot Compliance (EPRC, [1], [2]) concept presents a new level of compliance control paving the way for robotics into active and dynamic environments. Plastic compliance refers to evasive robot motions induced by interactions with active environments, after which the robot is not pushed back to the initial point of contact in contrast to elastic compliance. Providing increased softness particular to active environments that are recognized in an energy-based fashion, especially safety and robustness of cobots are enhanced by EPRC. Furthermore, cooperation with humans and other robots is enhanced via EPRC by putting robots into a subordinate role as required by current norms and standards. Thereby, each degree-of-freedom (DoF) is treated separately such that the robots can take leading roles in specific DoF and following roles in others. Aside these capabilities, the EPRC simplifies interaction with articulated objects increasing robustness in execution of complex tasks. At the same time, the EPRC does not require force sensors, disturbance observers or models of the environment and does not restrict interaction forces in contrast to comparable concepts. The EPRC was validated in experiments involving robot cooperation [1], [4] controlled from the ISS in a teleoperation setting [3] and in a healthcare assistance scenario.

Keywords: Rehabilitation Robotics, Social HRI, Medical Robots and Systems
Abstract: Physiotherapy is important in children after trauma or surgery in order to prevent pneumonia, cardiovas- cular impairment and/or constraint joints as well as in therapy of children with depressive disorder. However, experienced pediatric physiotherapists are scarce and expensive. In the mobirobot project we employ NAO humanoid robots that adopt typical movements and motivation tasks to support physiotherapists in mobilizing children on the pediatric wards and in an ambulatory setting. Robots are especially suited for this task because children are generally positive about robots, and their embodiment is beneficial for demonstrating movement exercises and for exerting influence on motivation and compliance. Their transfer, however, hinges on their integration into clinical and ambulatory everyday practice and routines, and their acceptance, the compliance, and the commitment of children, parents, and medical caregivers. In the mobirobot project, we evaluate their efficiency through validated instruments and scoring systems. This video showcases the work and impressions of working with mobirobot in the clinical and ambulatory settings.

Keywords: Disassembly, Recognition, AI-Enabled Robotics
Abstract: Efficient and sustainable battery disassembly is a key challenge in the circular economy of electromobility. Traditional disassembly processes are manual, time-consuming, and pose significant safety risks. di.monta at IGMR | RWTH Aachen University develops intelligent robotic software skills and modular end-of-arm tooling to enable automated and safe battery disassembly. On dismantling module is AI-driven screw detection and removal, utilizing YOLOv8-based computer vision to process image and depth data in real time. Our system integrates Intel RealSense depth cameras, a six-axis Neura Lara 8 robot, and specially designed end-effectors to ensure high precision and efficiency. Our solution achieves 96% precision in screw detection, with an 87% success rate in automated screw removal. The system operates with ±0.5 mm positional accuracy despite depth deviations of up to ±1 cm in the data and achieves an average processing time of 12 seconds per screw removal. It reliably detects screws even in cluttered environments with occlusions. This video demonstrates how our technology enhances automated circular economy and resource conservation, achieving up to 90% more efficient disassembly times than manual processes. di.monta is more than just automation - it is a game-changer for the future of battery recycling.

Keywords: Perception for Grasping and Manipulation, Computer Vision for Automation, Perception-Action Coupling
Abstract: Humans possess the remarkable capacity to recognize the 3D structure and orientation of completely novel objects. Contrary to this, most few-shot deep learning models typically require large amount of training samples to achieve a comparable ability. In this work, we introduce a Hybrid Neuro-Graphics model that uses neural foundation models in combination with a differentiable renderer to perform zero-shot pose estimation without any additional training data. Our model estimates the 6D poses of previously unseen objects from a single RGBD image and a bounding box by solving a series of constraint-based optimization problems. Moreover, we are able to recover scene parameters such as material properties, meshes, and lighting conditions, and use these predictions to perform zero-shot grasping. This ability to perform zero-shot pose estimation and grasping is crucial for robotic systems, as it enables them to interact with and manipulate novel objects in dynamic environments without the need for extensive retraining or pre-programmed knowledge.

Keywords: Machine Learning for Robot Control, Data Sets for Robot Learning, Software Tools for Benchmarking and Reproducibility
Abstract: The intersection of robotics and artificial intelli- gence led to a profound paradigm shift in Robot Learning. Robots have the capacity to replicate human actions and also dynamically adapt, innovate, and excel across a spectrum of tasks. However, the heterogeneity in the deployment of robot platforms and software frameworks poses considerable chal- lenges in terms of systematic testing and comparative analyses. Additionally, the data scarcity of especially force controlled robot manipulation is still restraining the development of advanced foundation models. A reference platform with default software stack can help to increase comparability, reducing development time and collect a large amount of tactile robot manipulation data. To address on this problem, we developed a Parallel and Distributed Robot AI (PD.RAI) framework, comprising a scalable ensemble of Robot Learning Units (RLUs), a global database, and the Robot Cluster Intelligence (RoCI). Each RLU is endowed with robot arms, cameras, and local computational units to autonomously engage in planning, control, and local machine learning of tactile manipulation skills. The RoCI system oversees the learning process and schedules the RLUs tasks. To show the functionality of the system, two black-box optimization algorithms are compared within the robot skill learning domain. An experiment with 24 different optimization tasks is conducted in parallel. The algorithms are incorporated into the same existing default modules acting as a reference environment. This allows for a realistic comparison without sacrificing diversity of possible configurations and testing environments.

Keywords: Soft Robot Applications, Marine Robotics, Biomimetics
Abstract: Navigating on the water surface allows swimming robots to aid in environmental monitoring or research in ecology. However, the surface is often highly unstructured, cluttered with plants, animals, and debris, which requires agility. We report cm-scale (45 mm length), fast (5 cm/s), and maneuverable (195 °/s) soft swimming robots with autonomous operation. Propulsion comes from traveling waves that are excited along undulating pectoral fins. The robot consists of a flat sub-mm-thin locomotion module (actuators and fins), power supply and control electronics. The locomotion modules monolithically integrate undulating fins with soft electrohydraulic actuators that operate at voltages below 500 V, at low power (<35 mW), at high bandwidth (>100 Hz), and that are durable (>750,000 actuation cycles). Modular design strategies extend locomotion capabilities beyond forward swimming and turning to include backward and sideways swimming, offering maneuverability similar to aerial quadcopters. We demonstrate autonomous operation facilitated by the integration of sensors, energy supply, power conversion, and control. The swimming robots circumnavigate obstacles, swim through narrow spaces, and push away heavy objects, enabling practical tasks through high thrust generation. This agile robot combines advances in swimming performance, robustness, and functionality, and will inspire future miniaturized vehicles operating in aquatic environments.

Keywords: Virtual Reality and Interfaces, Acceptability and Trust, Reinforcement Learning
Abstract: High-performance robot policies achieve high performance through neural networks trained via reinforcement learning. However, their black-box nature poses challenges for human-robot interaction, particularly regarding trust and behavior anticipation. To address this issue, we present a novel interactive virtual reality (VR) interface that visualizes explainable AI (XAI) outputs in an intuitive and user-friendly manner.

The interface allows users to interact with a reinforcement learning-based robot navigation policy, visualizing its LiDAR-based perception and decision-making process in real time. Key innovations include the projection of XAI attribution scores onto scene semantics, highlighting influential objects visually based on their impact on the robot's decisions while it is moving. Users can set start and goal positions, trigger the policy, and observe how different obstacle configurations affect the robot's behavior.

Our system transforms typically abstract XAI data into intuitive visual feedback, hypothesizing that this approach will enhance users' understanding of robot perception, improve behavior anticipation, and foster well-calibrated trust in both the robot and its policy. An ongoing user study aims to validate these hypotheses by examining how the interface impacts users' ability to understand and predict robot behavior in various scenarios.

Keywords: Agent-Based Systems, Distributed Robot Systems, Planning, Scheduling and Coordination
Abstract: Customized production demands flexible hybrid manufacturing systems with seamless subsystem interaction. Structured agent-based architectures, as e.g. proposed by VDI/VDE 2653, enable systems’ scalability but face challenges in information exchange due to unclear semantics and nonuniform communication interfaces. Existing standards like OPC UA and DDS can provide technical solutions to standardize communication but require semantic design for the exchanged information. Furthermore, a single technology may not suit heterogeneous environments like KI.Fabrik (AI.Factory) integrating production, logistics, robotics, and digital twins. This study introduces a semantically enriched message model for exchanging a priori knowledge, experiential learning, and fault-related data among agents. Designed based on agent communication requirements, the model ensures interoperability across diverse system architectures and adapts to various communication protocols like OPC UA. This approach was validated through two use cases, where five autonomous agents dynamically manage resource allocation. The first use case represents normal operation, while the second addresses resource shortages, requiring the coordination agent to find alternatives through agent collaboration. A real-time visualization tool was developed to monitor agent communication. Additionally, latency measurements showed the feasibility of extending the approach to industrial environments. Future work includes a deeper analysis of communication latency, network constraints in digital twins, and human-agent integration.