Keywords: Task Planning, Semantic Scene Understanding, Formal Methods in Robotics and Automation
Abstract: The Concept Hierarchy is a knowledge modeling framework for representing geometric, semantic, and dynamic scene elements in household environments. It stores necessary information for autonomous systems to plan and reason in indoor environments and serves typical applications thereof: environment modeling, action modeling and recognition, and task planning. Its hierarchical structure supports generalization and knowledge transfer to new entities thanks to the non-monotonic modeling of concept properties. We define tasks, actions, skills, and affordances that enable human-understandable and -explainable household applications. We validate the framework for action and skill recognition in human manipulation and for verifying a correct task execution in the environment.

Keywords: Manipulation Planning, Motion Control, Motion and Path Planning
Abstract: Energy efficiency and motion smoothness are essential in trajectory planning for high-degree-of-freedom robots to ensure optimal performance and reduce mechanical wear. This paper presents a novel framework integrating sinusoidal trajectory generation with velocity scaling to minimize energy consumption while maintaining motion accuracy and smoothness. The framework is evaluated using a physics-based simulation environment with metrics such as energy consumption, motion smoothness, and trajectory accuracy. Results indicate significant energy savings and smooth transitions, demonstrating the framework's effectiveness for precision-based applications. Future work includes real-time trajectory adjustments and enhanced energy models.

Keywords: Model Learning for Control, Intelligent and Flexible Manufacturing, Industrial Robots
Abstract: As a result of Industry 4.0 and the integration cyber physical systems and smart manufacturing, several industries moved to fully automating their processes. However, some industrial robots still need to be teleoperated to perform some tasks. These processes rely on human expertise due to one or more factors: 1) lack of suitable sensors, 2) lack of proper task representation, or 3) demand for precise and fine-grained manipulation skills. This paper visits an industrial robot task, which is highly challenging to automate and thus requires human teleoperation. We propose a framework for assisting operators in such tasks.

Keywords: Embodied Cognitive Science, Perception-Action Coupling, Semantic Scene Understanding
Abstract: Cognitive robotics requires robots to reason about tasks, objects, and environments in a way that closely resembles human flexibility and adaptability. This paper introduces Actionable Knowledge Graphs (AKGs) as a transformative approach to achieving this goal. By extending traditional knowledge graphs to include task-specific action knowledge and semantic digital twins encoding environment knowledge, AKGs enable robots to integrate environmental context into reasoning and dynamically adapt to novel situations. This work explores the construction and integration of AKGs into cognitive robotic architectures, highlighting their ability to resolve task ambiguity, infer motion parameters, and enhance the perception-action loop. Applications in flexible task execution, such as meal preparation, demonstrate the potential of AKGs to advance cognitive robotics in dynamic environments.

Keywords: Visual Tracking, Visual Servoing, Visual Learning
Abstract: We describe a visual SLAM system that integrates hand-crafted and learnable features, leveraging their complementary strengths for improved keypoint detection and descriptor representation. A pre-processing module is designed for feature processing, along with an efficient matching strategy for data association. Experiments conducted on the EuRoC-MAV and TUM-RGBD benchmarks demonstrate superior performance compared to related work, while maintaining comparable computational speed.

Keywords: Autonomous Vehicle Navigation, Sensor-based Control, Cooperating Robots
Abstract: The launch and recovery of an autonomous underwater vehicle (AUV) is a critical phase for the operation of these vessels and one of the most likely points of failure. One of the major risk factors are weather conditions, which may even prevent the process altogether. We address this issue with a novel launch and recovery system (LARS) comprised of a highly maneuverable AUV-like docking station tethered to a supply vessel. While this allows the mitigation of unfavorable weather conditions, it requires a higher degree of autonomous capabilities. In this paper, we present a novel method for facilitating physical docking underwater between two autonomous underwater vehicles, while both of them are in motion and with tight tolerances. This is achieved by blending between control laws based on a custom distance metric.

Keywords: Object Detection, Segmentation and Categorization, Semantic Scene Understanding, Deep Learning Methods
Abstract: Obtaining detailed information about instance geometry in 3D scenes is crucial for applications in robotics, augmented reality, and virtual reality. However, conventional surface reconstruction methods often suffer from coarse reconstruction quality, noise, occlusions, and non-watertight meshes. These limitations pose significant challenges for simulation environments, where physics-based interactions and collision handling require watertight 3D models. To address this issue, we propose leveraging Scan-to-CAD methods. Our approach utilizes publicly available 3D CAD models to detect, retrieve, align, and deform objects in real-world 3D scenes, resulting in a watertight, CAD-based representation. This representation facilitates seamless integration into physics-based simulations, simplifying and enhancing a wide range of robotics, augmented reality, and virtual reality applications. In this extended abstract, we outline the core criteria for effective Scan-to-CAD and present the strategies we developed to advance the state of the art. Our primary focus is on improving point cloud analysis through rotation-equivariant networks, which enhance representation learning, feature extraction, point cloud registration, and retrieval processes. By addressing these challenges, our work provides a robust foundation for high-quality 3D scene reconstruction and its application in dynamic, physics-based, simulation-driven environments.

Keywords: Marine Robotics, Motion Control
Abstract: During the First and Second World Wars, more than 1.6 million tons of unexploded ordnance (UXO) was dumped in the North and Baltic Seas, posing significant safety risks for the environment due to the gradual release of toxic chemicals. The project CleanSeas addresses this problem by studying AI algorithms for autonomous UXO disposal. In this work, we give an overview of the control algorithms used by the autonomous underwater vehicle (AUV) Cuttlefish to accomplish autonomous UXO inspection and manipulation tasks. First, we describe an orbital motion controller to do 360° scans of UXO. Furthermore, we also describe an algorithm called "Incremental Nonlinear Dynamic Inversion", which allows to control marine vehicles without the need for hydrodynamic models by leveraging high-fidelity measurements of acceleration and thrust. Not relying on hydrodynamic models is especially beneficial for systems with complex geometry such as Cuttlefish, since highly nonlinear and turbulent effects are hard to model.

Keywords: Reinforcement Learning, Planning under Uncertainty, Model Learning for Control
Abstract: Model-based approaches have shown great potential in planning and control for robotics. When a learned dynamics model is used, error accumulation during model-based rollouts negatively impacts long-horizon planning. We explore this problem from a Model-Based Reinforcement Learning (MBRL) perspective and present Infoprop, a novel rollout mechanism designed to generate trajectories with probabilistic models that closely resemble ground truth. Additionally, Infoprop quantifies the accumulation of model uncertainty throughout the trajectory. This accumulated uncertainty quantification enables rollouts to terminate once the uncertainty surpasses a threshold. Integrating Infoprop into state-of-the-art MBRL, we report substantial improvements in prediction horizon and accuracy.

Keywords: Mobile Manipulation, Search and Rescue Robots, Product Design, Development and Prototyping
Abstract: This paper presents an approach to design a topology-optimized modular robotic arm for integration with a legged robot. The primary objective is to enhance the robot's versatility and efficiency in performing complex tasks across various environments. Traditional robotic arms often face weight, structural integrity, and adaptability challenges. To address these issues, we employ topology optimization techniques to minimize material usage while maintaining structural performance, resulting in a lightweight yet robust design. The modular nature of the robotic arm allows for easy customization and maintenance, enabling the legged robot to switch between different end-effectors and configurations based on the task requirements. The design process incorporates advanced simulation tools to predict performance under various load conditions, ensuring reliability and functionality. Additionally, integrating sensors and control systems is discussed, providing insights into the arm's responsiveness and precision in dynamic scenarios. The experimental results demonstrate significant load capacity, maneuverability, and energy efficiency improvements compared to conventional designs.

Keywords: Social HRI, Education Robotics, Human-Centered Robotics
Abstract: Social educational robots hold promise for enhancing learning outcomes but face challenges in modeling and inducing engagement aligned with learning goals. Our recent research introduced Productive Engagement (PE), a goal-oriented construct focused on engagement that supports learning. Using multimodal student behavioral data from a collaborative educational human-human-robot interaction activity with 300 students over 3 years, we proposed innovative methodologies to model, characterize, and assess PE in real-time. Our results demonstrate that robots informed by PE statistically influence students’ engagement and improve learning outcomes, marking a paradigm shift in Human-Robot Interaction (HRI) design principles that are applicable across various closed-domain goal-oriented HRI applications. Building on these findings, we are currently modeling Productive Rapport (PR) between children with neurodevelopmental disorders (NDDs) and their caregivers aiming to enhance early communication skills via a robot mediator.

Keywords: Social HRI, Education Robotics, Human-Centered Robotics
Abstract: This research explores the use of Large Language Models (LLMs) for free conversational role-play in Socially Assistive Robots (SARs), addressing gaps in prior studies that focused on non-social contexts or Wizard-of-Oz setups. Recognizing the potential of role-playing to support cognitive, social, and emotional development across diverse users, we developed an interface allowing users or caregivers to create immersive role-play scenarios with LLM-powered agents. Preliminary evaluations using a sociolinguistic framework reveals that while LLMs can maintain and shift roles, challenges remain in playing some roles better than others, and in recognizing a user-driven break from pretend play. Additionally, we proposed a novel language assessment framework that extends beyond traditional metrics of factuality and faithfulness, assessing non-referential language functions critical for contextually rich interactions, such as in role-playing, that are not always factually correct. Initial results demonstrate promise but highlight the need for optimization techniques, larger-scale studies in real-world settings, and interdisciplinary methods to advance LLM-powered conversational role-playing companions.

Keywords: Motion and Path Planning, Formal Methods in Robotics and Automation, Intelligent Transportation Systems
Abstract: Just like human drivers, automated vehicles must explicitly comply with driving laws to obtain a legal license to operate. This requires safe maneuvering and strict adherence to all road traffic regulations under all conditions. However, existing motion planners often fail to guarantee safety or only consider a limited subset of traffic rules. To address this issue, we propose a framework that repairs planned trajectories monitored as violating driving laws -- particularly traffic rules -- serving as a seamless bridge between the planning and control modules. We integrate our approach into a research vehicle and validate it through real-world tests. The evaluation results confirm that our approach is computationally efficient, enforcing legal safety on automated vehicles online so that they meet the requirements for a driving license.

Keywords: Swarm Robotics, Collision Avoidance, Motion and Path Planning
Abstract: Distributed Model Predictive Control (DMPC) is a promising approach for safely controlling swarms of unmanned aerial vehicles (UAVs) by leveraging the scalability of distributed systems and the adaptability of dynamic control. However, computational and communication challenges have prevented existing DMPC approaches from being implemented on truly distributed hardware with wireless communication. In this extended abstract, we present DMPC-Swarm, a novel swarm architecture that, for the first time, enables DMPC on distributed hardware with wireless communication. Our system combines three key elements: an efficient, low-power wireless mesh communication protocol; distributed, event-triggered off-board computing that allows the use of nano quadcopters; and a novel DMPC algorithm that provably avoids UAV collisions even under communication message loss. By implementing DMPC-Swarm on a swarm of up to 16 nano quadcopters, we demonstrate its capability to safely and dynamically control quadcopter swarms. A video of our experiments is available at http://tiny.cc/DMPCSwarm.

Keywords: Computer Vision for Automation, Service Robotics, Incremental Learning
Abstract: This work presents a robust stereo-vision-based person re-identification system tailored for Human-Robot Interaction (HRI) in dynamic and open public environments.The method overcomes challenges like near proximity between humans and visually similar appearances, by leveraging an explicit spatial ambiguity model, to make associations based either on visual features or spatial proximity. The proposed system integrates a part-based feature extractor for visual association and spatial association via a Kalman filter, both based on YOLO v8 detections, to handle spatial ambiguity and occlusions. A prototype calibration mechanism ensures continuous updates of the target representation for resilience against appearance changes and out-of-FOV scenarios. Evaluation on a public dataset demonstrates superior performance compared to existing MOT and SOT algorithms, achieving success rates above 95% in selected sequences.

Keywords: Learning from Demonstration, Telerobotics and Teleoperation
Abstract: This work tackles the challenges of learning from teleoperated demonstrations, where communication impairments, such as delays, degrade demonstration quality and hinder learning outcomes. We present a network-aware framework that extends Hidden Semi-Markov Models (HSMM) and Task-Parameterized HSMM (TP-HSMM) with reliability-weighted learning. By utilizing quality metrics derived from teleoperation transparency, the framework dynamically adjusts the weighting of demonstration segments to counteract the impact of noisy or suboptimal data. The proposed approach is validated on a rubber band placement task using variable impedance control, achieving up to a 63% improvement in task success rates.

Keywords: Constrained Motion Planning
Abstract: We introduce a new type of soft constraint for HQP-based motion generation that preserves the desired direction of a task motion if the task constraint cannot be satisfied perfectly. In contrast to the usually employed soft quadratic constraint, this ensures straight-line trajectories in Cartesian space. Furthermore, we report on an interactive editor that facilitates the configuration of complex hierarchical tasks utilizing RViz interactive markers and enabling rapid exploration of various parameters and solver settings.

Keywords: Robotics and Automation in Agriculture and Forestry, AI-Enabled Robotics, Long term Interaction
Abstract: Collecting large datasets in robotics is important for various research tasks and naturally requires reliable long-term data acquisition. Therefore, this depends on an adequately designed, reliably functional robotic platform and support infrastructure. This work sketches such a platform and infrastructure and demonstrates its application in an exemplary use-case in agricultural plant monitoring.

Keywords: Agricultural Automation, AI-Enabled Robotics, Mechanism Design
Abstract: Strawberries, known for their fragile nature, require careful handling to prevent bruising and spoilage. Traditionally, harvesting is performed by skilled workers, however, persistent labor shortages and rising production costs have created a demand for automation. This extended abstract presents an autonomous agricultural robot designed for the selective harvesting of strawberries in open field ridging cultivation. The system features a scalable main body, along with a passive suspension that maintains stability and traction on uneven terrain. It autonomously navigates the field and employs a multi-spectral computer vision approach to identify ripe strawberries. The harvesting apparatus includes a low-inertia, belt-driven manipulator for fast picking, and a soft robotic gripper with a suction cup to isolate individual fruits.

Keywords: Reinforcement Learning, Imitation Learning, Legged Robots
Abstract: Using Deep Reinforcement Learning (DRL) for robotics often results in brittle and undeployable learning outcomes. To push the agent towards the expected solution, we propose to use a prior obtained from an expert demonstration. At the example of quadruped locomotion, we show that our latent action priors combined with established style rewards lead to the expected behavior, even achieving above expert demonstration level of performance. The latent action priors can be obtained from a micro-dataset. Further, our action priors substantially improve the performance on new tasks, such as walking at different speeds. Videos and the code are available at https://tinyurl.com/GRCpaper.

Keywords: Learning from Demonstration, AI-Enabled Robotics, Representation Learning
Abstract: In this work, we present a new method of learning a representation and a control policy from expert demonstra- tions. We approximate geodesics on the demonstration manifold based on the position of fitted prototypes and their connectivity representing the data topology. A first experiment is also conducted and the results are reported.

Keywords: Continual Learning, Neurorobotics, Distributed Robot Systems
Abstract: In order to ease collaborative robots (cobots) integration to the industrial welding process, the CORINNE project leverages neuromorphic sensing and computing to perform online, on chip learning of visual action recognition in a federated, semi-supervised fashion.

Keywords: Localization, Data Sets for SLAM
Abstract: Accurate localization is crucial for the autonomous operation of mobile robots. Specifically for indoor scenarios, localization algorithms typically rely on a previously generated map. However, many real-world sites like warehouses or healthcare environments violate the underlying assumption that the robot's surroundings are mainly static. In this paper, we introduce a new dataset that enables evaluating and comparing indoor localization methods in complex and changing real-world scenarios. While several datasets for indoor scenes exist, only a few combine the long-term localization aspect of repeatedly revisiting the same environment under varying conditions with precise ground truth over multiple rooms. Our dataset comprises various sequences recorded with a wheeled robot covering an office environment. We provide data from two 2D LiDARs, multiple consumer-grade RGB-D cameras, and the robot's wheel odometry. By densely placing fiducial markers on every room ceiling, we also provide accurate pose information within a single global frame for the whole environment. This enables benchmarking localization approaches over complete trajectories in challenging real-world scenarios.

Keywords: Learning from Demonstration, Reinforcement Learning, Imitation Learning
Abstract: Learning robot skills from a limited number of hu- man demonstrations can enable more flexibility in automation and service robotics. To this end, Offline Reinforcement Learn- ing is a promising approach as it aims to learn a meaningful policy from a fixed dataset without online interactions. However, one challenge in this approach is distributional shifts between the learned and behavior policies that lead to overestimating the value of out-of-distribution actions. Several methods exist that introduce conservatism into the value estimation to mitigate this problem. However, they are often evaluated on expert datasets in simulation using proprioceptive data. In this work, we explore their applicability on a simulated and a real robot pouring task. We thus provide insights into the performance of such methods under imperfect conditions such as noisy human demonstrations or image noise.

Keywords: Software Architecture for Robotic and Automation, AI-Enabled Robotics, Distributed Robot Systems
Abstract: The integration of AI into Cyber-Physical Production Systems (CPPS) is reshaping modern manufacturing, yet it raises critical safety and security concerns. This paper introduces a Model-Based Adaptive Data Basin framework that ensures controlled data access and constrained AI behavior within predefined, model-driven boundaries. The framework dynamically adapts to evolving factory environments, enhancing safety and operational reliability. Designed for multi-vendor and collaborative settings, the approach offers a robust foundation for secure AI integration, addressing the challenges of transparency and risk mitigation in AI-driven industrial systems.

Keywords: Simulation and Animation, Soft Sensors and Actuators
Abstract: Robotic research increasingly relies on modular software frameworks like ROS to enable flexible and extensible development. Simulation plays a critical role in this domain, offering opportunities for rapid prototyping, safe experimentation, and parallelized testing without hardware constraints. However, achieving a balance between simulation reliability and computational efficiency remains a challenge. In this work, we present two significant contributions to robotic simulation: (1) MuJoCo ROS, a novel simulation framework that combines the constraint-based MuJoCo engine, renowned for its multi-joint dynamics and contact-rich interaction capabilities, with seamless ROS integration, and (2) the use of this modular framework to incorporate advanced hydroelastic contact models for tactile sensor simulation. Together, these innovations advance the field of robotics by providing scalable, accurate, and extensible tools for diverse applications.

Keywords: Methods and Tools for Robot System Design, Cognitive Modeling, Social HRI
Abstract: Large Language Models (LLMs) and Vision-Language Models (VLMs) have the potential to significantly advance the development and use of cognitive architectures for social robots. Using a combined system consisting of an Adaptive Control of Thought-Rational (ACT-R) model and a humanoid social robot, I investigate how content from the declarative memory of the ACT-R model can be retrieved using real-world data obtained by the robot via an LLM or VLM, processed according to the procedural memory of the cognitive model and returned to the robot as instructions for action. In addition, real-world data captured by the robot can be stored as memory chunks in the cognitive model. This opens up possibilities for using human-like judgment and decision-making capabilities like instance-based learning and intuitive decision-making inherent in cognitive architectures with social robots and practically offers opportunities of augmenting the prompt for LLM-driven utterances with content from declarative memory thus keeping them more contextually relevant.

Keywords: Deep Learning Methods, Haptics and Haptic Interfaces
Abstract: The Tactile Internet aims to enable remote physical interactions, enhancing applications like teleoperation and bringing the human sense of touch to distant environments. This requires the encoding of haptic information to ensure efficient utilization of network resources. This extended abstract explores the requirements for compressing vibrotactile signals and provides an overview of the current state-of-the-art codecs designed for this purpose. Potential future directions in this field are also outlined.

Keywords: Medical Robots and Systems, Surgical Robotics: Steerable Catheters/Needles, Flexible Robotics
Abstract: Ultrahigh field (UHF) magnetic actuation using static background magnetic field of magnetic resonance imaging (MRI) scanners shows great promise in minimally invasive medical operations due to their high torque wireless actuation capabilities in remote parts of the patient’s body and inherent integration into medical imaging. However, the static nature of the MRI magnetic field limits the dexterity of UHF-actuated continuum robots. In this work, we proposed a UHF-actuated magnetic concentric-tube robotic catheter for increased dexterity. Our design is composed of a Teflon catheter with a permanent magnetic tip, a pre-curved core, and an MRI-compatible concentric gear actuator with piezo motors. We present a quasistatic kinematic model of the magnetic concentric-tube catheter in free space and demonstrate the workspace in a 7 T preclinical MRI scanner. Our design will pave the way for a high degree-of-freedom dexterous UHF-actuated catheter designs in MRI-guided minimally invasive medical operations.

Keywords: Learning from Demonstration, Imitation Learning
Abstract: MP are a well-established method for generating modular robot trajectories. This work introduces FA-ProDMP, a novel approach that adds force dimensions to ProDMP. The FA-ProDMP enables real-time adaptation of the desired trajectory based on measured forces by capturing position-force correlations across multiple demonstrations.

It supports multiple force axes, making it suitable for both Cartesian and Joint space control. To evaluate FA-ProDMP, we present POEMPEL, a modular, 3D-printed task suite inspired by Lego Technic pins. The base task suite presents a peg-in-hole task with adjustable parameters. Assembly instructions are available at https://intuitive-robots.github.io/poempel/

We show through experiments that FA-ProDMP outperforms other MP formulations on the POEMPEL setup and an electrical power plug insertion task by utilizing the position-force correlations of the recorded demonstrations. These results demonstrate FA-ProDMP's effectiveness in enhancing robotic performance in contact-rich manipulation tasks.

Keywords: Integrated Planning and Learning, Legged Robots, Machine Learning for Robot Control
Abstract: We present a novel hierarchical framework using reinforcement learning (RL) for learning robotic policies grounded in contacts. While policies trained via reinforcement learning have shown impressive results for locomotion, such results have still not emerged for complex loco-manipulation tasks. Using contact goals, which constitute contact locations and timings as subgoals for our low-level controller, our approach allows for more precise control of the end-effector interactions, while enabling generation of diverse interaction modes. We demonstrate the effectiveness of our method on a quadrupedal robot for executing multiple different gaits, such as trotting, bounding, pacing and jumping, using a single policy, and seamlessly transitioning between them. Our framework, taking first steps toward unifying locomotion and manipulation learning through contact, paves the way for more versatile and adaptable legged robots capable of complex loco-manipulation behaviours.

Keywords: Deep Learning in Grasping and Manipulation, Perception for Grasping and Manipulation, Probabilistic Inference
Abstract: We present vMF-Contact, a novel framework for learning hierarchical contact grasp representations through the probabilistic modeling of directional uncertainty using the von Mises–Fisher (vMF) distribution. In contrast to existing approaches that primarily address aleatoric uncertainty arising from inherent data noise, our method effectively captures epistemic uncertainty—critical for the recognition of out-of-distribution (OOD) objects—without relying on computationally intensive ensemble techniques, which are impractical for real-time deployment. Experimental evaluations in real-world settings demonstrate that our approach achieves a 39% improvement in overall clearance rate compared to baseline methods, underscoring its robustness and efficacy.

Keywords: Safety in HRI, Human-Robot Collaboration, Parallel Robots
Abstract: This extended abstract summarizes our research contributions on safe human-robot collaboration with parallel robots, which have a potential for safe and fast physical interaction due to their low moving masses. Methods for detection, classification, localization and reaction to collisions and clamping contacts with a planar parallel robot are described. Data-driven models for classification are designed, whose generalizability to unknown contact scenarios during training is facilitated by physically modeled features. Since only proprioceptive information is used, the methods can be transferred to already assembled parallel robots without installing new hardware components.

Keywords: Natural Dialog for HRI, Social HRI, Learning from Experience
Abstract: A key aspect of intuitive and natural Human-Robot-Interaction (HRI) is personalization, i.e. adapting the robot's behavior to align with the current user's needs and preferences. While recent approaches leverage foundation models to enable natural HRI, personalization remains an underexplored challenge. In this extended abstract, we propose a system that incrementally learns personalized humanoid robot behavior from natural-language interactions. Specifically, we build upon our existing dialog system that uses a large language model (LLM) to generate high-level code to steer the robot's behavior, and compare two approaches for personalization: (i) explicit user profiles by storing structured user-specific facts and allowing the LLM to determine when to access or update these memories, and (ii) incremental interactive learning by extending our existing learning mechanism to handle multiple user-specific interaction memories over time. We present preliminary experimental results in simulation and further demonstrate our system on the real-world humanoid robot ARMAR-DE. Our findings highlight that personalization remains an important but complex and non-trivial challenge for future research.

Keywords: Humanoid and Bipedal Locomotion, Legged Robots, Biomimetics
Abstract: The structure of the human foot, including the mobile arch and toe joint, plays a key role in efficient and adaptive locomotion. However, whether replicating these biomechanical features in robotic and prosthetic feet provides functional benefits remains an open question. Here, we present a newly developed bio-inspired, modular foot design for the EcoWalker bipedal robot. The modular foot enables a systematic comparison between rigid and mobile foot models, including a mobile arch and toe joint with passive elastic ligaments. The foot can be used in three configurations: a rigid foot, a foot with midtarsal joint motion but no toe joint, and a fully coupled three-segment foot. Elastic ligaments realize the passive dynamics and coupling of the joints. The foot prototype was manufactured using 3D printing, with encoders measuring joint motion. Future validation through static loading tests and robotic gait trials will assess the impact of these designs on bipedal locomotion. This work contributes to advancing robotic feet toward more efficient, human-like gait performance, with potential applications in both, robotics and prosthetics.

Keywords: Robot Companions, Social HRI, Human-Centered Robotics
Abstract: The objective of this study is to investigate whether the use of a privacy-preserving camera filter and robot transparency can encourage individuals to self-disclose in human-robot conversations. A 2 × 2 between-subject study was conducted with 28 participants, who engaged in a conversation with the PixelBot robot. Results suggest that self-disclosure depth is negatively affected by the personality trait ”extroverted” and not affected by the physical filter nor the robot transparency.

Keywords: Localization, Sensor Fusion, Learning Categories and Concepts
Abstract: This short paper presents research findings on two learning-based methods for quantifying measurement uncertainties in global navigation satellite systems. We investigate two learning strategies: offline learning for outlier prediction and online learning for noise distribution approximation, specifically applied to GNSS pseudorange observations. To develop and evaluate these learning methods, we introduce a novel multisensor state estimator based on continuous-time factor graph optimization. This estimator accurately and robustly estimates trajectory from multiple sensor inputs, critical for deriving GNSS measurement residuals used to train the uncertainty models. We validate the proposed learning-based models using real-world sensor data collected in diverse urban environments. Experimental results demonstrate that both models effectively handle GNSS outliers and improve state estimation performance. Furthermore, we provide insightful discussions to motivate future research toward developing a federated framework for robust vehicle localization in challenging environments.

Keywords: Perception for Grasping and Manipulation, Deep Learning in Grasping and Manipulation, Sensor Fusion
Abstract: Scientists have been researching robots for decades in order to enable them to manipulate objects like humans. But it is still a challenge to manipulate deformable objects dexterously. In the automotive industry, there has been significant interest in robotized wire harness assembly. At our institute, a series of depth image based tracking solutions for shape-variant cables and complex branched wire harnesses have been implemented. For more complex wire harnesses, graph- based topology matching is enabled using feature extraction. In the future, the learning-based graph neural network approach will be investigated for its potential in modelling the dynamics of wire harnesses. This approach demonstrates notable real-time capabilities. The fusion of visual sensing and force information is set to become the next breakthrough in robot sensing solutions, offering enhanced reliability and functionality.

Keywords: Domestic Robotics, AI-Enabled Robotics
Abstract: This paper introduces VLM-Vac, a novel framework aimed at improving the autonomy of smart robot vacuum cleaners. Our method combines the zero-shot object detection capabilities of a Vision-Language Model (VLM) with a Knowledge Distillation (KD) approach. The VLM enables the robot to classify objects into actionable categories--either to avoid or to suck---across various environments. However, frequent queries to the VLM are computationally expensive and impractical for real-world applications. To mitigate this, we employ KD to progressively transfer the VLM's essential knowledge to a compact, more efficient model. Real-world experiments confirm that this smaller model continuously learns from the VLM while significantly reducing the number of queries needed over time. Furthermore, we address continual learning in dynamic home settings by introducing a novel experience replay technique based on language-guided sampling. Our results indicate that this approach not only enhances energy efficiency but also outperforms traditional vision-based clustering methods, particularly in detecting small objects within complex backgrounds.

Keywords: Imitation Learning, AI-Enabled Robotics, Acceptability and Trust
Abstract: This paper introduces state-graphs as a new interpretable state representation for Imitation Learning (IL). Goal-Conditioned Imitation Learning (GCIL) requires efficient state representations for effective task execution. One common state representation are state-vectors, but state-graphs offer a more dynamic alternative. This paper leverages Graph Embeddings for EXplainable Imitation Learning (GraphEXIL). GraphEXIL shows the efficiency of graphs in handling complex, everyday tasks. State-graphs significantly outperform state-vectors in heterogeneous task groupings. They handle arbitrary numbers of objects and complex environments more efficiently, compared to state-vectors. Additionally, they are inherently explainable, enhancing interpretability for non-experts.

Keywords: Industrial Robots, Engineering for Robotic Systems, Hardware-Software Integration in Robotics
Abstract: When companies want to use mobile robots (MRs), they face many challenges. The implementation MR systems is associated with a high level of effort in terms of integration and adjustments to the infrastructure. In addition, the solutions are usually tailored to the customer’s infrastructure and needs. Approaches to standardize these processes already exist, but there are still gaps. The Machine to X (M2X) project tries to close these gaps and standardizes interfaces between MRs and their environment together with partners from industry in order to significantly reduce implementation costs and spread MRs more quickly. With our work, we use the success story of M2X to show how communication and collaboration between research and industry needs to be realized.

Keywords: Medical Robots and Systems, Compliance and Impedance Control
Abstract: This work discusses a previously published robotic platform for intraoperative ultrasound tissue scanning. The system integrates a novel approach to robotic interaction control. In contrast to other common force control approaches, high performance control at low forces and the difficulty of choosing an appropriate force profile is not required. The collaborative nature of the platform is reflected in its capability to execute autonomous, teleoperated or hands-on tissue scanning. Further assessment and experiments are necessary in proving the effectiveness of the proposed platform, to take a step towards versatile, collaborative, and safe robotic intraoperative ultrasound scanning across medical domains.

Keywords: Deep Learning Methods, Deep Learning in Grasping and Manipulation, AI-Based Methods
Abstract: Grasping objects successfully from a single-view camera is crucial in many manipulation tasks. An approach to solve this problem is to leverage simulation to create large datasets of pairs of objects and grasp poses, and then learn a conditional generative model that can be prompted quickly during deployment. However, the grasp pose data is highly multimodal since there are several ways to grasp an object. In this work, we learn a grasp generative model with diffusion models to sample candidate grasp poses given a partial point cloud of an object. We show in real-world experiments that our approach can grasp several objects from raw depth images with 90% success rate and benchmark it against several baselines.

Keywords: Machine Learning for Robot Control, Model Learning for Control, Reinforcement Learning
Abstract: NA.

Keywords: Multi-Contact Whole-Body Motion Planning and Control, Imitation Learning, Manipulation Planning
Abstract: Locomotion and manipulation are difficult tasks in robotics, as they involve a long-horizon decision-making problem with a combination of discrete and continuous decision variables. While simple end-to-end imitation and reinforcement learning have shown promise in the past few years, they generally struggle with problems that need reasoning over a long horizon. In this paper, we propose a structured approach to learning long-horizon locomotion problems. Our approach combines Monte-Carlo tree search (MCTS) to efficiently search over discrete decision variables (e.g., which surface to make contact with) and gradient-based trajectory optimization for checking the feasibility of the candidate contact plans. Since the whole process is still time-consuming and cannot be done for real-time control, we propose to leverage imitation learning (in particular diffusion models) to learn a policy that can reactively generate new feasible contact sequences. We tested our whole pipeline on quadrupedal locomotion on stepping stones in a simulated environment. Future works involve real-world experiment and extending this framework to solve loco-manipulation tasks.

Keywords: Safety in HRI, Transfer Learning, Parallel Robots
Abstract: Supervised-learning methods hold significant potential for contact classification in human-robot collaboration (HRC). However, they require a sufficient amount of data for effective training. In the context of HRC, data generation can be hazardous, costly, and labor-intensive. Motivated by these challenges, this extended abstract provides insights into the benefits of transfer learning for contact-body classification in HRC with parallel robots (PRs). Differently scaled PR simulations serve as source domains, while the classifier aims to generalize to a real PR as the target domain. To address the sim-to-real gap, an approach with domain randomization and adaptation is employed. Three transfer-learning tasks are investigated, with results demonstrating that the sim-to-real gap is reduced. Notably, the best performance is achieved by combining simulated and real-world data, surpassing the performance of models trained exclusively on real-world data.

Keywords: AI-Based Methods, Reinforcement Learning, Learning from Experience
Abstract: In this paper, we propose a concept used federated reinforcement learning (FRL) framework designed to facilitate the transfer of learned robot skills, such as peg-in hole insertion tasks. This framework enables new robots to acquire task-specific skills through a shared global model while maintaining the privacy of their sensors and environmental data. We introduce a novel FRL framework to overcome the challenges associated with skill transfer in robotic systems.

Keywords: AI-Enabled Robotics, Cognitive Control Architectures
Abstract: Robot cognitive architectures increasingly emphasize the importance of memory as an active component, bridging the gap between semantic understanding and sensorimotor experiences. It not only manages the flow of information between different processes but also provides essential services: extracting semantic meaning from sensorimotor data, parameterizing symbolic plans with actionable parameters, and predicting the outcomes of actions. Crucially, such memory systems must not only acquire and retain information but also selectively forget - a process that prevents information overload, reduces computational overhead, and helps maintain relevant, up-to-date knowledge for decision-making. This extended abstract describes our work on implementing forgetting mechanisms in the deep episodic memory of a humanoid robot. We report on the results of various experiments to describe the methods used to evaluate the effects of forgetting mechanisms on robot task performance over extended periods in real-world experiments.

Keywords: Representation Learning, Neurorobotics, Machine Learning for Robot Control
Abstract: This work explores the latent representation of trajectories resulting from concrete actions generated by a Central Pattern Generator (CPG) model. Using a variational Autoencoder (VAE) with an additional regularization loss to ensure smoothness in the reconstructed trajectories, we could map the 2D planar motion into a low-dimensional latent space. The latent representation is then used for a downstream task of predicting the concrete action parameters of the CPG model.

Keywords: Motion and Path Planning, Grasping, Manipulation Planning
Abstract: Batch planning is increasingly necessary to quickly produce diverse and high-quality motion plans for downstream learning applications, such as distillation and imitation learning. This paper presents Global Tensor Motion Planning (GTMP)---a sampling-based motion planning algorithm comprising only tensor operations. We introduce a novel discretization structure represented as a random multipartite graph, enabling efficient vectorized sampling, collision checking, and search. We provide a theoretical investigation showing that GTMP exhibits probabilistic completeness while supporting modern GPU/TPU. Additionally, by incorporating smooth structures into the multipartite graph, GTMP directly plans smooth splines without requiring gradient-based optimization. Experiments on lidar-scanned occupancy maps and the MotionBenchMarker dataset demonstrate GTMP's computation efficiency in batch planning compared to baselines, underscoring GTMP's potential as a robust, scalable planner for diverse applications and large-scale robot learning tasks.

Keywords: Agent-Based Systems, AI-Enabled Robotics, Autonomous Agents
Abstract: The field of robotics is currently experiencing a new boom, driven in part by the hype surrounding generative AI. This paper provides an insight into how Large Language Models (LLMs) can enhance communication between robots and humans, making interactions more interactive and intuitive. Additionally, it explores how a robotic system can autonomously solve abstract tasks with the help of AI-Agents, focusing specifically on construction-related challenges as examples building structures with Lego stones.

Keywords: Performance Evaluation and Benchmarking, Mapping
Abstract: The generation of maps of the environment is one of the tasks for which mobile sensing systems, such as robots, are utilized. These maps are created using various perception sensors, such as LiDARs, in conjunction with the pose information of the system. Determining the accuracy of these maps is challenging due to multiple influencing factors, including pose estimation and system calibration, which impact the final map. We propose a method to benchmark the 3D mapping accuracy of mobile sensing systems using a freely accessible test environment with highly accurate reference data. We evaluate the accuracy of various parameters derived from the generated 3D map of the test environment, which are relevant for real-world applications. Our approach can assess mapping accuracy under different conditions, such as changing environmental settings, and provides insights into correlations primarily arising from pose estimation.

Keywords: Imitation Learning, Kinematics, Learning from Demonstration
Abstract: Imitation is a major aspect of intelligence behavior, which provides a versatile and rapid mechanism to transfer motor skills from one intelligent agent (e.g., human, animal, or robot) to another. One major challenge in imitation learning is the correspondence problem: establishing corresponding states and actions between an expert demonstrating the task and a learner trying to imitate the task when their embodiments are dissimilar (morphology, dynamics, degrees of freedom, etc.). Many existing approaches circumvent the correspondence problem by directly providing demonstrations of the robot learner, for example, in kinesthetic teaching or teleoperation. These methods require robot-specific proprioception, which is not always available. In this study, we investigate the correspondence problem between dissimilar embodiments. In particular, we propose a correspondence divergence between embodiments and derive an imitation policy via the proposed divergence in the reactive motion generation or inverse reinforcement learning settings. We conducted the first set of experiments with increasing complexity of embodiments, showing that the approach is well suited for identifying morphology correspondence in robot imitation.

Keywords: Computer Vision for Transportation, Aerial Systems: Applications, Intelligent Transportation Systems
Abstract: This study addresses the challenge of transportation unit (TU) identification in inland port terminals, which frequently change storage area configurations and handle different unit types such as containers and semi-trailers. Conventional methods rely on fixed infrastructure or fixed structure, often lack flexibility and come at a high cost, making them unsuitable for smaller ports. Current computer vision solutions are limited in handling TUs with varying scales and orientations, particularly in mobile perspectives.

We propose a solution using unmanned aerial vehicles (UAVs) and sensor-equipped reach stackers to enhance terminal monitoring at the Alberthafen port in Dresden. Our three-stage identification pipeline includes TU detection, text field detection, and text recognition, utilising lightweight models suited for edge devices. YOLOv5-S is deployed for initial TU detection to reduce the search space for small text, followed by DBNet++ for text detection in diverse conditions and finally, the TrOCR model recognise the ID codes with low character error rates. These models were fine-tuned on the TRUDI dataset, containing 25,941 labelled instances across multiple TU classes.

Our approach integrates aerial and ground-level data collection, enhancing accuracy. UAVs, such as the DJI Mavic Pro 3 and Mini 2, capture high-resolution images to create orthophotos for a georeferenced terminal overview. Sensor-equipped reach stackers complement this by capturing ground-level data, ensuring precise tracking of the TUs.

Initial results show promising performance; the TU detection model achieves a mean average precision (mAP) of 0.71, and DBNet++ yields an F1-score of 0.85, while TrOCR's character error rate (CER) is notably low at 0.02. These results demonstrate robustness compared to existing methods, laying the way open for further automation and optimisation in port operations.

Keywords: Reactive and Sensor-Based Planning, Mapping, Semantic Scene Understanding
Abstract: Metric-semantic mapping methods’ reliance on overconfident semantic predictions and noisy depth data, coupled with the inherent problem of 2D to 3D assignment lead to poor quality semantic map generation. Firstly, we propose our evidential semantic mapping framework that integrates evidential depth and semantic predictions from monocular RGB images using a novel evidential multi-task learning network. Secondly, the semantic layer uses Bayesian Kernel Inference for spatial regression of input measurements to produce smoother maps. Thirdly, we present an active semantic view planning methods that utilizes the evidential metric-semantic uncertainties for generating high information gain viewpoints. A semantic collision checker utilizes the evidential semantic information to produce semantically safe trajectories for environment traversal in cluttered environments like glasshouses and households. Our active semantic view planning method combined with the evidential semantic mapping framework will lead to calibrated map generation that enables robust semantic scene understanding.

Keywords: Reinforcement Learning
Abstract: This work introduces Transformer-based Off-Policy Episodic Reinforcement Learning (TOP-ERL), a novel algorithm that enables off-policy updates in an ERL framework. In ERL, policies predict entire action trajectories over multiple time steps instead of single per-step actions. These trajectories are typically parameterized by trajectory generators such as Movement Primitives (MP), allowing for smooth and efficient exploration over long horizons while capturing temporal correlations. However, ERL methods are often constrained to on-policy frameworks due to the difficulty of evaluating state-action values for action sequences, limiting their sample efficiency and preventing the use of more efficient off-policy architectures. TOP-ERL addresses this shortcoming by segmenting long action sequences and estimating the state-action values for each segment using a transformer-based critic architecture alongside an nstep return estimation. These contributions result in efficient and stable training that is reflected in the empirical results conducted on sophisticated robot learning environments. TOPERL significantly outperforms state-of-the-art RL methods. Thorough ablation studies additionally show the impact of key design choices on the model performance

Keywords: Human-Robot Collaboration, Assembly, Robust/Adaptive Control
Abstract: This extended abstract presents an innovative research direction and achievements to date. We aim to create robots capable of adapting to their human co-worker during industrial assembly tasks that have a high product variety. Our approach combines the following research areas: 1) monitoring and predicting human behavior, 2) assembly sequence planning, and 3) robot adaptation. Preliminary results are available for all three areas. We briefly summarize the current state of each research area and present the missing components that we plan to address the next two years.

Keywords: Medical Robots and Systems, Machine Learning for Robot Control, Motion Control
Abstract: Electromagnetic navigation enables precise remote control of magnetic objects, offering advanced medical applications such as targeted drug delivery and minimally invasive surgeries. Commonly controlled devices include magnetic catheters, guidewires, and micro- to nanorobots. In this video, we shift our focus from clinical applications to showcase the dynamic capabilities of electromagnetic navigation systems by stabilizing a 3D inverted pendulum. Through this work, we aim to advance the development of dynamic control algorithms for electromagnetic navigation, addressing the current reliance of the research area on quasi-static modeling and feedforward control. Although the inverted pendulum itself has no direct clinical relevance, it serves as an ideal platform for developing and testing novel magnetic control algorithms. By adopting a model-based design approach, we successfully stabilize pendulums ranging from 20 to 40 cm in length. Additionally, we incorporate an iterative learning control scheme that enables precise tracking of non-equilibrium trajectories while maintaining the stability of the inverted pendulum. Our work showcases the potential of dynamic models and advanced control algorithms to significantly enhance the capabilities of electromagnetic navigation systems.

Keywords: Human-Centered Robotics, Social HRI, AI-Based Methods
Abstract: This video presents the design of the EchoSync study, focusing on the evaluation of different voice design paradigms for android robots. The study aims to compare the perceived likability and anthropomorphism of interaction scenarios featuring distinct voice design approaches. Specifically, the video demonstrates two design paradigms: (1) design congruence, where the robot speaks with a crowd-sourced voice that matches its appearance, and (2) user similarity, where the robot uses a cloned voice of the study participants. Additionally, a baseline condition was included, featuring a voice unrelated to either the robot's appearance or the participants. Preliminary analysis of the collected data reveals that both design paradigms — design congruence and user similarity — are perceived as significantly more likable than the baseline condition. The study investigates voice cloning as a possible approach to user-centered voice design and compares two major voice design paradigms for humanoid robots, providing insights that could guide future applications.

Keywords: Aerial Systems: Applications, Agent-Based Systems, Software, Middleware and Programming Environments
Abstract: The growing demand for autonomous drone collectives to perform adaptive missions and diverse applications requires frequent reconfiguration while ensuring safety during design and runtime. Compositional modeling and verification frameworks offer a promising foundation for studying self-adaptive drone collectives, enhancing interoperability and coordination. However, existing methods often struggle with true plug-and-play modularity, where system extensions preserve existing requirements and behaviors without causing model fragmentation or requiring extensive re-verification.

This video introduces the foundational steps of a compositional modeling platform for drone swarm coordination, based on the bigraphical reactive systems framework. This framework defines drone behaviors in terms of movement and interaction, constrained spatially and temporally. Temporal rules capture the progression of distributed system states, while spatial dimensions represent state spaces and causal relationships. By ensuring causal closure, the framework enables safe behavior composition, provided cyber-physical consistency is maintained.

Bridging the gap between formal methods and real-world applications, the proposed approach delivers a low-code framework for system design, verification, and implementation, leveraging meta-modeling and integrated toolchains. A simple case study demonstrates how program composition simplifies multi-drone formations, showcasing its adaptability and correctness-by-design methodology.

Keywords: Aerial Systems: Mechanics and Control, Biologically-Inspired Robots
Abstract: Several bird species have the ability to hover and maneuver efficiently in vertical winds by leveraging the energy of the surrounding airflow. Thruster-powered UAVs, such as quadcopters, offer exceptional maneuverability but are limited by high energy consumption and an inability to efficiently leverage environmental energy. In contrast, fixed-wing aircraft and gliders excel in energy efficiency and can utilize environmental energy but lack the ability to hover or perform agile maneuvers. We introduce Floaty, a soaring robot designed to bridge the gap between fixed-wing and thruster-powered UAVs. Floaty leverages vertical wind for passive, energy-efficient flight while maintaining the ability to hover and execute agile maneuvers through the use of large adjustable flaps that dynamically control its aerodynamic profile. Its design is inspired from the soaring capabilities of birds, achieving position and attitude control without sacrificing energy efficiency. Performance evaluations in a custom-built wind tunnel demonstrate Floaty's ability to hover, track attitude and position. Furthermore, we analyze the influence of design and scale on its efficiency and dynamics. Our results demonstrate the feasibility of a bio-inspired, energy-efficient approach to soaring in dynamic wind conditions, opening up new avenues for the design of future aerial robotic systems.

Keywords: Telerobotics and Teleoperation, Human-Robot Collaboration, Haptics and Haptic Interfaces
Abstract: This video illustrates telerobotics over 5G mobile networks, demonstrating its application in cross-city tasks and emphasizing the essential role of reliable, high-quality communication between the leader and follower systems. High-fidelity teleoperation over long distances is often hampered by unavoidable communication network issues, such as latency, jitter, and packet loss. However, the rapid advancement of network technologies, such as 5G and the anticipated emergence of 6G mobile networks, is driving significant progress in teleoperation by effectively mitigating challenges like latency, jitter, and packet loss. In this video, we present results from a collaboration between researchers in Munich and Prague, focusing on cross-city teleoperation enabled by next-generation network technologies. Building on our previous work, which involved the implementation of two distinct teleoperation testbeds, we validate their performance through real-world experimental evaluations in this study.

Keywords: Agricultural Automation
Abstract: Agriculture faces several challenges including climate change and biodiversity loss while, at the same time, the demand for food, feed, biofuels, and fiber is increasing. Sustainable intensification aims to increase productivity and input-use efficiency while enhancing the resilience of agricultural systems to adverse environmental conditions through improved management and technology. Recent advances in sensing, machine learning, modeling, and robotics offer opportunities for novel smart digital technologies to enable sustainable intensification. However, developing smart digital technologies and putting them into agricultural practice, requires closing major research gaps, related in particular to (1) the utilization of multi-scale multi-sensor monitoring in space and time, (2) using artificial intelligence for linking process and data-driven methods, (3) improving decision making and intervention in plant production, and finally (4) modeling conditions and consequences of farmers acceptance. Closing these gaps requires an interdisciplinary approach. Here, we present a research agenda and steps forward to steer research efforts, highlighting research priorities, and identifying required interdisciplinary research collaboration. Following this agenda will leverage the full potential of smart digital technologies for sustainable crop production.

Keywords: Agricultural Automation
Abstract: On the path to sustainable crop production, the use of robotic systems could play a major role. While Unmanned Aerial Vehicles (UAVs) are increasingly used to monitor the health of agricultural fields with sensors such as RGB or spectral cameras and LiDAR, it is often still necessary to physically enter the field to conduct close-up inspections of individual plants or even plant organs, such as leaves, to detect diseases or nutrient deficiencies at early stages of growth.

The video demonstrates the integration of aerial and ground robotic systems to automate the plant inspection processes, thereby enhancing efficiency in field monitoring tasks. A UAV identifies areas or plants of interest from the air, and their coordinates are transferred to an Unmanned Ground Vehicle (UGV). The UGV then automatically navigates to the specified location, where a mounted robotic arm with five cameras captures close-up images with automatically optimized camera positions. The result is a high-resolution 3D reconstruction suitable for further plant analysis. This system integration allows us to 'zoom in' on any coordinate in the field with an accuracy of a few centimeters.

Keywords: Multi-Robot Systems, Motion and Path Planning, Motion Control
Abstract: Motion planning problems for physically-coupled multi-robot systems in cluttered environments are challenging due to their high dimensionality. Existing methods combining sampling-based planners with trajectory optimization produce suboptimal results and lack theoretical guarantees.

We propose pc-dbCBS, a kinodynamic motion planner for high-dimensional physically-coupled systems. Our method extends discontinuity-bounded Conflict-Based Search (db-CBS) to the physically-coupled systems domain. Our hybrid approach uses a discrete search over motion primitives that are computed for individual robots offline and allows bounded violations for the physical coupling constraints. The resulting solution is transformed to a different minimal-state representation that is then used by a trajectory optimization, enforcing physical coupling constraints implicitly. By repeating these steps iteratively, the resulting algorithm becomes probabilistically complete and asymptotically optimal.

We demonstrate on a benchmark with 25 problems in simulation and 6 problems on real robots that our method is generalizable across different robot types, namely cable-suspended payload transport using multirotors and differential-drive robots connected via rigid rods. Our approach outperforms the state-of-the-art by solving more instances and producing solutions that are twice as fast with significant lower computational effort.

Keywords: Legged Robots, Reinforcement Learning
Abstract: The field of legged robotics is still missing a single learning framework that can control different embodiments -such as quadruped, humanoids, and hexapods - simultaneously and transfer, zero or few-shot, to unseen robot embodiments. To close this gap, we introduce URMA, the Unified Robot Morphology Architecture. Our framework brings the end-to-end Multi-Task Reinforcement Learning approach to the realm of legged robots, enabling the learned policy to control any type of robot morphology. We show that URMA can learn a locomotion policy on multiple embodiments that can be transferred to unseen robots in simulation and the real world.

Keywords: Computer Vision for Automation, Deep Learning Methods, Visual Learning
Abstract: Effective bone fracture reconstruction is crucial for patient outcomes, yet existing methods struggle with spatial misalignment and predictive accuracy. This work leverages machine learning for surgical planning, treating bone reconstruction as an image in-painting task. We employ Variational Autoencoders (VAE) to generate estimated healthy bone structures from fractured CT scans.

Our key contributions include (i) a novel adaptation of Spatial Transformer Networks (STN) for 3D spatial registration, enabling robust alignment of CT fragments, (ii) a comparative evaluation of autoencoder architectures for modeling bone CTs, and (iii) an extension of these techniques to masked CT images, allowing predictive reconstruction of healthy bone structures.

Keywords: Marine Robotics, SLAM, Micro/Nano Robots
Abstract: Sonar is an important sensor for underwater map- ping and perception, which are central building blocks for au- tonomy and intelligent behaviors in marine robotics. We present here an overview of results from the project ”Unconstrained Synthetic Aperture Sonar (U-SAS)”. This includes Fourier Soft in 2D (FS2D) as a novel registration method for noisy sonar data and Synthetic Scan Formation (SSF) as an approach to enhance mapping with Mechanically Scanning Sonars (MSS). These methods enable robust odometry estimation, loop closures, and real-time underwater mapping using low-cost devices like the BlueROV2. We demonstrate their effectiveness through field trials, including the mapping of the flooded basement of the Submarine Bunker Valentin in Bremen, contributing to the digitization of cultural heritage.

Keywords: Imitation Learning, Machine Learning for Robot Control, Collision Avoidance
Abstract: Diffusion policies have recently become popular in robotics, but their ability to handle unseen and dynamic conditions with novel constraints not represented in the training data is limited. To address this limitation, we propose diffusion predictive control with constraints (DPCC), an algorithm for diffusion-based control with explicit state and action constraints. By using constraint tightening and incorporating model-based projections into the denoising process of a trained trajectory diffusion model, DPCC can generate constraint-satisfying, dynamically feasible, and goal-reaching trajectories for predictive control. As shown through simulations of a robot manipulator, DPCC outperforms existing methods in satisfying novel constraints while maintaining task performance.

Keywords: Underactuated Robots, Wheeled Robots, AI-Based Methods
Abstract: Developing learning-based control (LBC) algorithms for robots with fast, unstable dynamics and diverse morphologies is a significant challenge. A major hurdle is the absence of versatile robotic platforms that capture these complexities while allowing for automatic and robust experimentation. We present the Mini Wheelbot, an underactuated unicycle robot that has many favorable characteristics in this regard. We mention recent research involving the Mini Wheelbot and describe the currently ongoing effort in building a general-purpose testbed and benchmark using the Mini Wheelbot for developing novel LBC algorithms.

Keywords: AI-Enabled Robotics, Imitation Learning
Abstract: This work introduces Variational Diffusion Distillation (VDD), a novel method that distills denoising diffusion policies into Mixtures of Experts (MoE) through variational inference. Diffusion Models are the current state-of-the-art in imitation learning due to their exceptional ability to accurately learn and represent complex, multi-modal distributions. However, diffusion models come with some drawbacks, including the intractability of likelihoods and long inference times due to their iterative sampling process. The inference times, in particular, pose a significant challenge to real-time robot control. In contrast, MoEs effectively address the aforementioned issues while retaining the ability to represent complex distributions but are notoriously difficult to train. VDD is the first method that distills pre-trained diffusion models into MoE models, and hence, combines the expressiveness of Diffusion Models with the benefits of MoEs. VDD demonstrates across nine complex behavior learning tasks, that it is able to: i) accurately distill complex distributions learned by the diffusion model, ii) outperform existing state-of-the-art distillation methods, and iii) surpass conventional methods for training MoE.

Keywords: AI-Enabled Robotics, Machine Learning for Robot Control, AI-Based Methods
Abstract: Accurately modeling friction in robotic joints remains challenging. Traditional model-based approaches are complex and inflexible, while data-driven methods lack interpretability and trust. This work proposes symbolic regression (SR) as a promising alternative. SR generates interpretable formulas, similar to model-based approaches, while adapting to diverse scenarios. We apply SR to data from a DLR KUKA LWR IV+ robot, demonstrating its ability to achieve higher accuracy than with model-based methods featuring comparable complexity of the formulas.

Keywords: Mapping, Perception for Grasping and Manipulation, Deep Learning in Grasping and Manipulation
Abstract: Searching for objects in cluttered environments requires selecting efficient viewpoints and manipulation actions to remove occlusions and reduce uncertainty in object locations, shapes, and categories. In this work, we address the problem of manipulation-enhanced semantic mapping, where a robot has to efficiently reason about objects in a cluttered shelf. Although Partially Observable Markov Decision Processes~(POMDPs) are standard for decision-making under uncertainty, representing unstructured interactive worlds remains challenging in this formalism, due to its state space complexity. To tackle this difficulty, we define a POMDP whose belief is summarized by a metric-semantic grid map and propose a novel framework that uses neural networks to perform map-space belief updates to reason efficiently and simultaneously about object geometries, locations, categories, occlusions, and manipulation physics. We call these networks Calibrated Neural Accelerated Belief Update (CNABU) networks and show that they can generalize to novel scenarios and transfer well sim-to-real in zero-shot fashion.

Keywords: Machine Learning for Robot Control, Reinforcement Learning, Assembly
Abstract: Although extensive research has been conducted, robotic assembly tasks requiring high precision and contact-rich compliant interaction remain a challenge for manufacturing automation. Guiding the learning process has long been considered a way to improve the sample efficiency of learning-based approaches. We propose a novel guided Reinforcement Learning approach, Adaptive Manipulation Primitives (AMP), which integrates compliant manipulation behavior modeling via Manipulation Primitive Nets with Deep Reinforcement Learning. We demonstrate how to efficiently learn search strategies to handle position uncertainties in a demanding gear assembly benchmark task (shaft-bore tolerance: 18,textmu m) in the real world. Our proposed approach allows for a seamless spectrum between fully learned robot behavior and entirely model-based control, enabling adaptive and efficient solutions tailored to different automation requirements.

Keywords: Localization, AI-Based Methods
Abstract: Localization in dynamic indoor environments is a long-standing and challenging problem in robotics. Traditional methods relying on a static map struggle when significant environmental changes occur as they cannot distinguish between permanent and temporary landmarks. Building Information Models (BIMs) can aid localization in dynamic scenes since they offer a structured description of static building elements and their spatial relationships. We propose a Monte-Carlo localization method using a vision-language model (VLM) and a BIM as reference. Using RGB images captured by the robot, the VLM generates natural language descriptions of building elements in the scene. These descriptions are then compared with those extracted from BIM renderings, leveraging language similarity as the sensor model. We further report initial localization results in an office building.