Keywords: Humanoid and Bipedal Locomotion, Whole-Body Motion Planning and Control, Body Balancing
Abstract: "What can your humanoid robot do?” is probably the most commonly asked question that we, as roboticists, have to answer when interacting with the general public. Often, the question is framed in the familiar household or office setting, with implied expectations of robust locomotion on uneven and cluttered terrain, and compliant interaction with people, objects, and the environment. Moreover, the question implies the existence within the humanoid robot of a set of embodied loco-manipulation skills implemented by a motion planner, skills that are retrievable when given the corresponding commands. In this article, we formulate an answer to this question in the form of an efficient, modular, and extensible motion planner. We demonstrate its use with three challenging scenarios, designed to highlight both the robot's safe operation and its precise movement in unstructured environments. Additionally, we discuss key techniques derived from our experience in the practical implementation of torque-controlled humanoid robots.

Keywords: Legged Robots, Actuation and Joint Mechanisms, Reinforcement Learning
Abstract: This paper presents a 3-DOF hopping robot with a human-like lower-limb joint configuration and a flat foot, capable of performing dynamic and repetitive jumping motions. To achieve both high torque output and a large hollow shaft diameter for efficient cable routing, a compact 3K compound planetary gearbox was designed using mixed-integer nonlinear programming for gear tooth optimization. To meet performance requirements within the constrained joint geometry, all major components—including the actuator, motor driver, and communication interface—were custom-designed. The robot weighs 12.45 kg, including a dummy mass, and measures 840 mm in length when the knee joint is fully extended. A reinforcement learning-based controller was employed, and the robot’s performance was validated through hardware experiments, demonstrating stable and repetitive hopping motions in response to user inputs. These experimental results indicate that the platform serves as a solid foundation for future bipedal robot development.

Keywords: Humanoid and Bipedal Locomotion, Legged Robots, Multi-Contact Whole-Body Motion Planning and Control
Abstract: This paper uses the capabilities of latent diffusion models (LDMs) to generate realistic RGB human-object interaction scenes to guide humanoid loco-manipulation planning. To do so, we extract from the generated images both the contact locations and robot configurations that are then used inside a whole-body trajectory optimization (TO) formulation to generate physically consistent trajectories for humanoids. We validate our full pipeline in simulation for different long-horizon loco-manipulation scenarios and perform an extensive analysis of the proposed contact and robot configuration extraction pipeline. Our results show that using the information extracted from LDMs, we can generate physically consistent trajectories that require long-horizon reasoning.

Keywords: Robot Safety, Collision Avoidance, Optimization and Optimal Control
Abstract: Autonomous navigation through unstructured and dynamically-changing environments is a complex task that continues to present many challenges for modern roboticists. In particular, legged robots typically possess manipulable asymmetric geometries which must be considered during safety-critical trajectory planning. This work proposes a predictive safety filter: a nonlinear model predictive control (MPC) algorithm for online trajectory generation with geometry-aware safety constraints based on control barrier functions (CBFs). Critically, our method leverages Poisson safety functions to numerically synthesize CBF constraints directly from perception data. We extend the theoretical framework for Poisson safety functions to incorporate temporal changes in the domain by reformulating the static Dirichlet problem for Poisson's equation as a parameterized moving boundary value problem. Furthermore, we employ Minkowski set operations to lift the domain into a configuration space that accounts for robot geometry. Finally, we implement our real-time predictive safety filter on humanoid and quadruped robots in various safety-critical scenarios. The results highlight the versatility of Poisson safety functions, as well as the benefit of CBF constrained model predictive safety-critical controllers.

Keywords: Humanoid Robot Systems, Art and Entertainment Robotics, Humanoid and Bipedal Locomotion
Abstract: Humanoid robots represent the cutting edge of robotics research, yet their potential in entertainment remains largely unexplored. Entertainment as a field prioritizes visuals and form, a principle that contrasts with the purely functional designs of most contemporary humanoid robots. Designing entertainment humanoid robots capable of fluid movement presents a number of unique challenges. In this paper, we present Kid Cosmo, a research platform designed for robust locomotion and life-like motion generation while imitating the look and mannerisms of its namesake character from Netflix's The Electric State. Kid Cosmo is a child-sized humanoid robot, standing 1.45 m tall and weighing 25 kg. It contains 28 degrees of freedom and primarily uses proprioceptive actuators, enabling torque-control walking and lifelike motion generation. Following worldwide showcases as part of the movie's press tour, we present the system architecture, challenges of a functional entertainment robot and unique solutions, and our initial findings on stability during simultaneous upper and lower body movement. We demonstrate the viability of performance-oriented humanoid robots that prioritize both character embodiment and technical functionality.

Keywords: Actuation and Joint Mechanisms, Wheeled Robots, Humanoid and Bipedal Locomotion
Abstract: The functions demanded of a robot vary with the environment. For a transformable robot that can switch among multiple configurations, the ability to adopt a wheeled mode—which yields highly efficient and stable locomotion—is crucial. The locations where wheels are mounted strongly influence both the robot's performance in wheel mode and the range of environments it can handle. Although wheel placement dictates the wheel-mode posture, that posture must remain reachable from the legged mode through a feasible transformation. This paper proposes a method for selecting wheel placements that enable transformation between a bipedal mode and tri-swerve mode, together with an ankle joint design that realizes those placements. First, we establish a systematic taxonomy of wheel configurations and, based on this evaluation, demonstrate the advantages of a steering ankle joint. Second, to achieve both leg-wheel transformation and steered wheeled locomotion, each ankle incorporates a wrap-free wire differential that provides a wide workspace. Experiments show that the proposed robot exploits the pitch and roll axes of its wire-differential steering ankles to transform smoothly and to perform both bipedal and tri-swerve motion.

Keywords: Legged Robots, Mechanism Design, Reinforcement Learning
Abstract: Robotic locomotion research typically draws from biologically inspired leg designs, yet many human-engineered settings can benefit from non-anthropomorphic forms. TARS3D translates the block-shaped “TARS” robot from Interstellar into a 0.25 m, 0.99 kg research platform with seven actuated degrees of freedom. The film shows two primary gaits: a bipedal-like walk and a high-speed rolling mode. For TARS3D, we build reduced-order models for each, derive closed-form limit-cycle conditions, and validate the predictions on hardware. Experiments confirm that the robot respects its ±150° hip limits, alternates left-right contacts without interference, and maintains an eight-step hybrid limit cycle in rolling mode. Because each telescopic leg provides four contact corners, the rolling gait is modeled as an eight-spoke double rimless wheel.

The robot’s telescopic leg redundancy implies a far richer gait repertoire than the two limit cycles treated analytically, so we used deep reinforcement learning in simulation to search the unexplored space. We observed that the learned policy can recover the analytic gaits under the right priors and also discover novel behaviors. Our findings show that TARS3D’s fiction-inspired bio-transcending morphology can realize multiple previously unexplored locomotion modes, and that further learning-driven search is likely to reveal more. This combination of analytic synthesis and reinforcement learning opens a promising pathway for multimodal robotics.

Keywords: Rehabilitation Robotics, Whole-Body Motion Planning and Control, Physical Human-Robot Interaction
Abstract: Transfer assistance places a significant physical burden on caregivers, making the introduction of transfer assistance robots essential to alleviate this load. For such robots to become practical, they must be capable of fully autonomous assistance without human intervention and perform care actions that do not impose physical strain on the care receivers. In this study, to perform assistance actions based on human caregiver movements, an end-effector elliptical trajectory synchronized with the care receiver’s movements is generated, and force control adapted to the receiver’s physical condition is implemented. Furthermore, quantitative evaluation metrics are proposed to assess the kinematic load on care receivers during sitting-up assistance. The assistive motions performed by a life-sized humanoid robot using the proposed method are evaluated and compared with assistive motions following different trajectories and human assistance in terms of kinematic load and subjective assessments of the care receiver.

Keywords: Dexterous Manipulation, Reinforcement Learning
Abstract: Humans naturally coordinate multiple fingers to perform a wide range of daily tasks in an ergonomically efficient manner. In contrast, robotic hands have yet to fully leverage such dexterity, often executing uncoordinated actions despite their multi-fingered structures. To bridge this gap, we introduce the Synergetic Dexterity Challenge: How can robotic hands effectively synergize their fingers to enable intuitive and efficient object manipulation, akin to human-level dexterity? To address this challenge, we propose SynPolDex, a novel bi-level pipeline that integrates high-level planning with low-level reinforcement learning to achieve coordinated finger control. Specifically, our method consists of two components. First, we employ an explicit chain-of-thought approach using a vision-language model for multi-turn reasoning and reflection, which generates a task-specific finger synergy plan by categorizing fingers into either dominant or non-dominant roles. Second, we acquire a low-level policy to execute motions based on the synergy plan. To incorporate the finger plan as prior knowledge, we reshape the reward function to encourage appropriate contact, and applies an advantage factorization technique to enhance policy learning. We evaluate SynPolDex across a variety of tasks and three different robot embodiments. Experimental results show that SynPolDex achieves higher success rates, faster convergence, and higher human-likeness score compared to baseline methods.

Keywords: Imitation Learning, Learning from Demonstration, Motion and Path Planning
Abstract: We propose an imitation learning framework that enables robots to acquire speed-adaptive manipulation skills from expert demonstrations exhibiting diverse motion velocities. Traditional imitation learning methods often assume temporally aligned, uniformly paced demonstrations, leading to policies constrained to a narrow range of execution speeds. However, real-world human demonstrations, particularly those collected in-the-wild without explicit instruction, naturally vary in motion speed. When trained on such data, conventional behavior cloning tends to produce averaged actions, often resulting in unreliable task execution. To address this limitation, we introduce a novel approach that conditions the policy on target motion velocities, enabling the generation of task-consistent actions at desired speeds. This capability is especially important for human-robot interaction, where the robot must flexibly adapt to human behavior. We evaluate the proposed method on object manipulation tasks, including grasp-and-place scenarios, in both simulated and real-world environments. Experimental results show that our method effectively learns speed-adaptive policies that generalize across a wide range of target velocities, outperforming standard imitation learning baselines.

Keywords: Multifingered Hands, In-Hand Manipulation, Dexterous Manipulation
Abstract: The rapid increase in the development of humanoid robots and customized manufacturing solutions has brought dexterous manipulation to the forefront of modern robotics. Over the past decade, several expensive dexterous hands have come to market, but advances in hardware design, particularly in servo motors and 3D printing, have recently facilitated an explosion of cheaper open-source hands. Most hands are anthropomorphic to allow use of standard human tools, and attempts to increase dexterity often sacrifice anthropomorphism. We introduce the open-source ISyHand (pronounced easy-hand), a highly dexterous, low-cost, easy-to-manufacture, on-joint servo-driven robot hand. Our hand uses off-the-shelf Dynamixel motors, fasteners, and 3D-printed parts, can be assembled within four hours, and has a total material cost of about 1,300 USD. The ISyHand’s unique articulated-palm design increases overall dexterity with only a modest sacrifice in anthropomorphism. To demonstrate the utility of the articulated palm, we use reinforcement learning in simulation to train the hand to perform a classical in-hand manipulation task: cube reorientation. Our novel, systematic experiments show that the simulated ISyHand outperforms the two most comparable hands in early training phases, that all three perform similarly well after policy convergence, and that the ISyHand significantly outperforms a fixed-palm version of its own design. Additionally, we deploy a policy trained on cube reorientation on the real hand, demonstrating its ability to perform real-world dexterous manipulation.

Keywords: Humanoid and Bipedal Locomotion, Robot Safety, Legged Robots
Abstract: Push recovery during locomotion will facilitate the deployment of humanoid robots in human-centered environments. In this paper, we present a unified framework for walking control and push recovery for humanoid robots, leveraging the arms for push recovery while dynamically walking. The key innovation is to use the environment, such as walls, to facilitate push recovery by combining Single Rigid Body model predictive control (SRB-MPC) with Hybrid Linear Inverted Pendulum (HLIP) dynamics to enable robust locomotion, push detection, and recovery by utilizing the robot's arms to brace against such walls and dynamically adjusting the desired contact forces and stepping patterns.

Extensive simulation results on a humanoid robot demonstrate improved perturbation rejection and tracking performance compared to HLIP alone, with the robot able to recover from pushes up to 100N for 0.2s while walking at commanded speeds up to 0.5m/s. Robustness is further validated in scenarios with angled walls and multi-directional pushes.

Keywords: Legged Robots, Reinforcement Learning, Humanoid and Bipedal Locomotion
Abstract: Loco-manipulation of quadrupedal robots has broadened robotic applications, but using legs as manipulators often compromises locomotion, while mounting arms complicates the system. To mitigate this issue, we introduce bipedalism for quadrupedal robots, thus freeing the front legs for versatile interactions with the environment. We propose a risk-adaptive distributional Reinforcement Learning (RL) framework designed for quadrupedal robots walking on their hind legs, balancing worst-case conservativeness with optimal performance in this inherently unstable task. During training, the adaptive risk preference is dynamically adjusted based on the uncertainty of the return, measured by the coefficient of variation of the estimated return distribution. Extensive experiments in simulation show our method's superior performance over baselines. Real-world deployment on a Unitree Go2 robot further demonstrates the versatility of our policy, enabling tasks like cart pushing, obstacle probing, and payload transport, while showcasing robustness against challenging dynamics and external disturbances.

Keywords: Aerial Systems: Mechanics and Control, Humanoid Robot Systems, Control Architectures and Programming
Abstract: We propose a novel Model Predictive Control (MPC) framework for a jet-powered flying humanoid robot. The controller is based on a linearised centroidal momentum model to represent the flight dynamics, augmented with a second-order nonlinear model to explicitly account for the slow and nonlinear dynamics of jet propulsion. A key contribution is the introduction of a multi-rate MPC formulation that handles the different actuation rates of the robot’s joints and jet engines while embedding the jet dynamics directly into the predictive model. We validated the framework using the jet-powered humanoid robot iRonCub, performing simulations in Mujoco; the simulation results demonstrate the robot’s ability to recover from external disturbances and perform stable, non-abrupt flight manoeuvres, validating the effectiveness of the proposed approach.

Keywords: Telerobotics and Teleoperation, Bimanual Manipulation, Dexterous Manipulation
Abstract: Teleoperation of humanoid robots refers to the real-time remote control of a robot by a human operator. An immersive approach to achieving this involves using cobots attached to the operator’s wrists, enabling both input capture and feedback delivery. The spatial configuration of these cobots relative to the human operator plays a crucial role in ensuring effective teleoperation. This paper presents novel human & robot workspace models, multi-stage optimisation processes, and dedicated evaluation metrics for determining the optimal placement strategy of one of the cobots (considering mirrored configurations) to enhance the quality and efficiency of humanoid teleoperation. The human workspace is represented as a maximally reachable mesh volume rather than as a set of reachable point clouds. To accelerate the optimisation for identifying the optimal placement, a volumetric intersection–based proposal stage is introduced. Two evaluation metrics—arm closeness and joint closeness index—are then employed to rank and extract the optimal placements from the optimisation results.

Keywords: Force and Tactile Sensing, Agricultural Automation, Sensor Fusion
Abstract: Contact-rich manipulation tasks in agriculture, such as pruning and harvesting, require robots to physically interact with tree structures to navigate through cluttered foliage. Identifying whether the robot is contacting rigid or soft materials is critical for the downstream manipulation policy to be safe, yet vision alone is often insufficient due to occlusion and limited viewpoint in this unstructured environment. To address this, we propose a multi-modal classification framework that fuses audio and visual inputs to identify the contact class: leaf, twig, trunk, or ambient. Our key insight is that contact-induced vibrations carry material-specific signals, making audio effective for detecting contact events and distinguishing broad material types, while visual features add complementary semantic cues that supports more fine-grained classification. We collect training data using a hand-held sensor probe and demonstrate zero-shot generalization to a robot-mounted probe embodiment, achieving an F1 score of 0.82. These results underscore the potential of audio-visual learning for manipulation in unstructured, contact-rich environments.

Keywords: Perception for Grasping and Manipulation, Perception-Action Coupling, Manipulation Planning
Abstract: Service robots operating in cluttered human environments such as homes, offices, and schools cannot rely on predefined object arrangements and must continuously update their semantic and spatial estimates while dealing with possible frequent rearrangements. Efficient and accurate mapping under such conditions demands selecting informative viewpoints and targeted manipulations to reduce occlusions and uncertainty. In this work, we present a manipulation-enhanced semantic mapping framework for occlusion-heavy shelf scenes that integrates evidential metric-semantic mapping with reinforcement-learning-based next-best view planning and targeted action selection. Our method thereby exploits uncertainty estimates from Dirichlet and Beta distributions in the map prediction networks to guide both active sensor placement and object manipulation, focusing on areas with high uncertainty and selecting actions with high expected information gain. Furthermore, we introduce an uncertainty-informed push strategy that targets occlusion-critical objects and generates minimally invasive actions to reveal hidden regions by reducing overall uncertainty in the scene. The experimental evaluation shows that our framework enables to accurately map cluttered scenes, while substantially reducing object displacement and achieving a 95% reduction in planning time compared to the state-of-the-art, thereby realizing real-world applicability.

Keywords: Motion and Path Planning, Imitation Learning, Sensorimotor Learning
Abstract: Deploying humanoid robots in everyday environments requires rapidly generating coordinated and collision-free motions for two arms. Despite recent advances in global motion planning and trajectory optimization, high-dimensional configuration spaces still pose a significant challenge and can cause long planning delays. Neural motion planning addresses this shortcoming by leveraging large planning problem datasets to train a motion planning policy. Such a model can then be used online or to seed global planners. In this work, we extend the neural motion planning approach to a dual-arm setting and thus generate motions in a 14-dimensional joint space. The proposed method is based on a transformer architecture and combines proprioceptive and geometric information in the form of point clouds to predict chunks of joint-space actions. In contrast to previous approaches, we use recent advances in imitation learning for fast point cloud tokenization. We show the performance of our model in simulation and provide insights for deploying it on a real robot. In our experiments, we use the model as a motion planner and for seeding a state-of-the-art trajectory optimizer. Our results show the advantage of transformer-based neural motion planning and its positive impact on the success rates of traditional solvers.

Keywords: Natural Dialog for HRI, Long term Interaction, Robot Companions
Abstract: Verbalization of robot experience, i.e., summarization of and question answering about a robot's past, is a crucial ability for improving human-robot interaction. Previous works applied rule-based systems or fine-tuned deep models to verbalize short (several-minute-long) streams of episodic data, limiting generalization and transferability. In our work, we apply large pretrained models to tackle this task with zero or few examples, and specifically focus on verbalizing life-long experiences. For this, we derive a tree-like data structure from episodic memory (EM), with lower levels representing raw perception and proprioception data, and higher levels abstracting events to natural language concepts. Given such a hierarchical representation built from the experience stream, we apply a large language model as an agent to interactively search the EM given a user's query, dynamically expanding (initially collapsed) tree nodes to find the relevant information. The approach keeps computational costs low even when scaling to months of robot experience data. We evaluate our method on simulated household robot data, human egocentric videos, and real-world robot recordings, demonstrating its flexibility and scalability.

Keywords: Deep Learning in Grasping and Manipulation, Semantic Scene Understanding, Transfer Learning
Abstract: Transferring skills between different objects remains one of the core challenges of open-world robot manipulation. Generalization needs to take into account the high-level structural differences between distinct objects while still maintaining similar low-level interaction control. In this paper, we propose an example-based zero-shot approach to skill transfer. Rather than treating skills as atomic, we decompose skills into a prioritized list of grounded task-axis (GTA) controllers. Each GTAC defines an adaptable controller, such as a position or force controller, along an axis. Importantly, the GTACs are grounded in object key points and axes, e.g., the relative position of a screw head or the axis of its shaft. Zero-shot transfer is thus achieved by finding semantically-similar grounding features on novel target objects. We achieve this example-based grounding of the skills through the use of foundation models, such as SD-DINO, that can detect semantically similar keypoints of objects. We evaluate our framework on real-robot experiments, including screwing, pouring, and spatula scraping tasks, and demonstrate robust and versatile controller transfer for each.

Keywords: Developmental Robotics, Cognitive Modeling, Sensorimotor Learning
Abstract: As robots become increasingly embedded in human environments, the ability to anticipate the outcomes of physical contact is crucial for enabling safe, adaptive, and socially intelligent behavior. Thus, learning to discriminate harmful sensory patterns from the benign ones will not only ensure physical safety during robot interaction, but may also lay the foundation for artificial empathy through mirroring the pain of others. To this end, this work develops a framework for tactile prediction through multimodal learning, emphasizing the integration of visual and tactile information in a common latent space. The ability to predict tactile sensations prior to contact allows a robot to avoid harmful outcomes as well as internalizing the tactile experience of others. We adapt the Deep Modality Blending Network (DMBN) as a foundational model for this task. Using demonstrations involving both gentle and noxious human touch, synchronized visual and tactile data are collected to train the model. After learning, the robot can generate temporal tactile activations from visual observations alone, anticipating sensory outcomes before physical contact occurs. Experiments on an upper-body humanoid robot show that it can predict painful stimuli and mirror tactile experiences observed in others. The key contributions of this study include: (1) the development of a predictive tactile perception framework using DMBNs, (2) the adaptation of this framework for modeling artificial pain that may be used as a basis for artificial empathy, and (3) empirical validation using real-world human-robot interaction scenarios.

Keywords: Tendon/Wire Mechanism, Mechanism Design, Actuation and Joint Mechanisms
Abstract: Operating in environments too harsh or inaccessible for humans is one of the critical roles expected of robots. However, such environments often pose risks to electronic components as well. To overcome this, various approaches have been developed, including autonomous mobile robots without electronics, hydraulic remotely actuated mobile robots, and long-reach robot arms driven by wires. Among these, electronics-free autonomous robots cannot make complex decisions, while hydraulically actuated mobile robots and wire-driven robot arms are used in harsh environments such as nuclear power plants. Mobile robots offer greater reach and obstacle avoidance than robot arms, and wire mechanisms offer broader environmental applicability than hydraulics. However, wire-driven systems have not been used for remote actuation of mobile robots. In this study, we propose a novel mechanism called Remote Wire Drive that enables remote actuation of mobile robots via wires. This mechanism is a series connection of decoupled joints, a mechanism used in wire-driven robot arms, adapted for power transmission. We experimentally validated its feasibility by actuating a wire-driven quadruped robot, which we also developed in this study, through Remote Wire Drive.

Keywords: Wearable Robotics, Human-Centered Robotics
Abstract: Accurately predicting where a person will place their foot during walking has practical value in applications that require close coordination between humans and machines, such as exoskeletons that adapt to a user's movement, or systems that detect and prevent trips and falls in real-world environments. Current methods often rely on complex models or offline analysis. In this paper, we present the use of probabilistic movement primitives (ProMPs) for predicting user step lengths in real time during walking on a treadmill. We used kinematic data acquired with an inertial measurement system to mimic data potentially gatherable from a wearable exoskeleton, avoiding the need for external motion capture. We evaluated the method with nine subjects walking on a treadmill. We show accurate prediction of user step length as early as 100 ms after movement onset during the foot's swing phase. This method could be extended and integrated with environmental monitoring systems to predict potential foot-obstacle collisions in real time.

Keywords: Manipulation Planning
Abstract: Autonomous manipulation of granular media, such as sand, is crucial for applications in construction, excavation, and additive manufacturing. However, shaping granular materials presents unique challenges due to their high-dimensional configuration space and complex dynamics, where traditional rule-based approaches struggle without extensive engineering efforts. Reinforcement learning (RL) offers a promising alternative by enabling agents to learn adaptive manipulation strategies through trial and error. In this work, we present an RL framework that enables a robotic arm with a cubic end-effector and a stereo camera to shape granular media into desired target structures. We show the importance of compact observations and concise reward formulations for the large configuration space, validating our design choices with an ablation study. Our results demonstrate the effectiveness of the proposed approach for the training of visual policies that manipulate granular media including their real-world deployment, outperforming two baseline approaches.

Keywords: Imitation Learning, Representation Learning, Learning from Demonstration
Abstract: Imitation learning offers an effective framework for enabling robots to acquire complex skills, but typically requires a large number of labeled demonstrations, making data collection costly. In contrast, large-scale unlabeled datasets containing diverse trajectories can be obtained more easily, motivating methods that leverage such data to enable learning from few demonstrations. A promising approach along this direction is to retrieve trajectories similar to demonstrations from unlabeled datasets. However, existing retrieval-based methods often rely on frame-level comparisons or fixed-length trajectory embeddings, limiting their ability to capture the temporal structure of behaviors and to generalize across different execution speeds. To address these limitations, this paper proposes a method that embeds variable-length trajectories using Self-Attention to capture sequence information and robustly retrieve semantically similar behaviors. Using a small number of demonstrations, relevant trajectories are extracted from unlabeled datasets and used alongside demonstrations to train the agent, thereby reducing data collection costs. Experiments on large-scale, diverse datasets demonstrate that the proposed method achieves higher retrieval accuracy and task success rates than existing retrieval-based imitation learning methods under realistic conditions.

Keywords: Mobile Manipulation, Service Robotics, Social HRI
Abstract: We present ANUBIS—a compact, compliant, and affordable humanoid mobile manipulation robot designed primarily for safe interaction and versatile task execution in everyday household environments. The platform integrates two lightweight, compliant 6 degree-of-freedom (DOF) arms onto a compact, cylinder-shaped omnidirectional mobile base, enabling agile navigation through standard doorways and tight indoor spaces. Torque-controlled quasi-direct-drive (QDD) actuators and a layered impedance control framework ensure inherent compliance, facilitating safe physical interactions with humans and household objects alike. Additionally, extensive use of off-the-shelf components and 3D-printed structures contributes to cost efficiency, keeping the total bill-of-materials near USD 11,200. Real-world experiments demonstrated autonomous vision-based manipulation and intuitive real-time bimanual teleoperation, validating the robot’s practical safety and versatility.

Keywords: Humanoid and Bipedal Locomotion, Reinforcement Learning, Body Balancing
Abstract: Bipedal locomotion is a key challenge in robotics, particularly for robots like Bolt, which have a point-foot design. This study explores the control of such underactuated robots using constrained reinforcement learning, addressing their inherent instability, lack of arms, and limited foot actuation. We present a methodology that leverages Constraints-as-Terminations and domain randomization techniques to enable efficient sim-to-real transfer. Through a series of qualitative and quantitative experiments, we demonstrate that our approach enables balance maintenance, velocity control, and effective slip and push recovery. Additionally, we analyze autonomy through metrics like the cost of transport and ground reaction force. Our method advances robust control strategies for point-foot bipedal robots, offering insights into broader locomotion.

Keywords: Multi-Contact Whole-Body Motion Planning and Control, Optimization and Optimal Control, Mobile Manipulation
Abstract: Over the years, robotic manipulation has primarily focused on end-effectors, an approach that severely limits a robot's ability to manipulate large and heavy objects. Humanoids, which are expected to operate in human environments, must acquire this skill to enhance their versatility and usefulness. In this regard, we present a whole-body multi-contact manipulation (WBMC) framework to handle large and heavy objects. To facilitate WBMC, we incorporate artificial skin patches distributed across the humanoid's upper body which are effectively utilized for contact detection and force sensing. The WBMC manipulation problem is formulated as an optimal control problem (OCP) within a model predictive control (MPC) framework, and three different types of dynamic motions are used to evaluate the controller's effectiveness. The proposed framework manages the entire manipulation process, including reaching, grasping, picking up, and manipulating. Furthermore, these motions are leveraged to develop a two-stage object inertial parameter estimation framework. The first stage estimates the object's mass and center of mass, while the second estimates its inertia. Both the manipulation and estimation frameworks are numerically evaluated using the TALOS humanoid and a rectangular box in simulation, and their respective results are presented and discussed.

Keywords: Human and Humanoid Motion Analysis and Synthesis, Bimanual Manipulation, Dual Arm Manipulation
Abstract: Bimanual manipulation tasks present significant challenges for robots due to the high dimensionality and complex coordination required between two arms. While humans perform such tasks effortlessly, transferring this capability to robots remains challenging. In this paper, we introduce a novel synergy-based representation for human bimanual manipulation that captures the characteristics of coordinated movements in a low-dimensional space, the synergy space. We train a variational autoencoder network on human motion data to learn synergies in bimanual manipulation tasks. We compare a Euclidean representation with a Riemannian representation in the latent space. Experimental results demonstrate that our synergy-based representation reduces dimensionality while improving both representation accuracy and motion smoothness. Specifically, our method achieves a 38% reduction in representation error compared to linear dimensionality reduction techniques, while simultaneously improving the smoothness of generated movements by 42%. We validate our approach through qualitative and quantitative analysis of reconstructed and newly generated bimanual movements, showing that the resulting motions preserve the essential characteristics of human manipulation.

Keywords: Dexterous Manipulation, Bimanual Manipulation, Deep Learning in Grasping and Manipulation
Abstract: The human hand is a remarkable feat of biology, with the ability to handle intricate tools with great precision and strength yet softly handle delicate objects. Robot hands attempting to emulate this have often fallen into one of two categories: soft or rigid. Soft hands, while compliant and yielding lack the precision and strength of human hands. Conversely, rigid hands are brittle to bumps and do not conform naturally to their environment. We call our solution LEAP Hand v2 Advanced, a dexterous, 3000, simple anthropomorphic hybrid rigid-soft hand that bridges this gap. First, it achieves a balance of human-hand-like softness and stiffness via a 3d printed soft exterior combined with a 3d printed internal bone structure. Next, LEAP Hand v2 Adv incorporates two powered articulations in the foldable palm: one spanning the four fingers and another near the thumb—mimicking the essential palm flexibility for human-like grasping. Lastly, LEAP Hand v2 Adv boasts a dexterous Metacarpophalangeal (MCP) kinematic structure, making it highly human-like, easy to assemble, and versatile. Through thorough real-world experiments, we show that LEAP Hand v2 Adv exceeds the capabilities of many existing robot hands for grasping, teleoperated control, and imitation learning. We release 3D printer files and assembly instructions for the dexterous hand research community at our website at https://v2-adv.leaphand.com/

Keywords: Grasping, Grippers and Other End-Effectors, Industrial Robots
Abstract: Industrial automation increasingly demands flexible and robust grasping solutions for objects of varied geometries and sizes. Conventional parallel-jaw grippers frequently struggle to achieve stable contact in complex industrial contexts, causing inefficiencies and downtime. To overcome such challenges, adaptive grippers characterized by multi-axis control and multi-contact capabilities are a commercially viable alternative. However, traditional volumetric grasping techniques catering to parallel-jaw grippers are not directly applicable due to the difference in contact mechanics and actuation constraints, necessitating the development of novel grasping algorithms for enabling flexible multi-point contact. In this paper, we propose a framework for learning volumetric grasp generation for large industrial objects using a sophisticated adaptive gripper comprising of eight independently movable axes and four vacuum cups. We develop a novel approach for generating grasp candidates by using inverse kinematics to search for suitable gripper configurations that align the point clouds of both the adaptive gripper and the target object surface to capture critical geometric features, yielding a robust set of multi-contact training examples. We then propose a neural rendering-based volumetric grasp detection approach to predict suitable grasp candidates for a multi-axis gripper based on the global and local geometric information of each end-effector. Our experiments validate the efficacy of our proposed grasp generation approach that achieves a promising grasp success rate of ~95% with different industrial objects. This highlights the effectiveness of combining adaptive multi-contact strategies with geometry-centric data generation.

Keywords: Actuation and Joint Mechanisms, Reinforcement Learning, Humanoid and Bipedal Locomotion
Abstract: Reinforcement learning (RL) has enabled advances in humanoid robot locomotion, yet most learning frameworks do not account for mechanical intelligence embedded in parallel actuation mechanisms due to limitations in simulator support for closed kinematic chains. This omission can lead to inaccurate motion modeling and suboptimal policies, particularly for robots with high actuation complexity. This paper presents general formulations and simulation methods for three types of parallel mechanisms: a differential pulley, a five-bar linkage, and a four-bar linkage, and trains a parallel-mechanism aware policy through an end-to-end curriculum RL framework for BRUCE, a kid-sized humanoid robot. Unlike prior approaches that rely on simplified serial approximations, we simulate all closed-chain constraints natively using GPU-accelerated MuJoCo (MJX), preserving the hardware's mechanical nonlinear properties during training. We benchmark our RL approach against a model predictive controller (MPC), demonstrating better surface generalization and performance in real-world zero-shot deployment. This work highlights the computational approaches and performance benefits of fully simulating parallel mechanisms in end-to-end learning pipelines for legged humanoids.

Keywords: Mechanism Design, Hydraulic/Pneumatic Actuators, Humanoid Robot Systems
Abstract: Wheeled biped robots’ intrinsic instability frequently limits their mobility and operational performance, reducing their capacity to adapt to different terrains and manipulate with stability. This research suggests a hydraulicdriven wheeled biped robot (WBR) strategy that can accomplish flexible transitions between wheeled and footed states in order to address this shortcoming. One of its striking characteristics is a servo valve-controlled transformable wheel that uses variable displacement control of six hydraulic cylinders to change the wheel’s shape from circular to triangular. Consequently, mode switching between wheels and feet is accomplished by the WBR robot. Furthermore, based on cylinder-valve-skeleton integration design and 3D printing technology, the WBR robot also features hydraulic-driven humanoid legs and dual 4-DOF humanoid arms, achieving highly integrated and lightweight development. Lastly, a wheel-foot switching control approach based on the transformable wheel mechanism is proposed in this study. Wheel movement and wheel-foot switching studies were used to confirm the hydraulic-driven wheeled biped robot system’s dependability and efficiency.

Keywords: Humanoid Robot Systems, Biomimetics, Dexterous Manipulation
Abstract: Robotic hands often struggle to replicate the dexterity and stability of human grasping, particularly in tasks requiring precise thumb opposition or passive compliance. Although most designs approximate the complex trapeziometacarpal (TMC) joint with simplified 2-degree-of-freedom mechanisms, they rarely optimize axis placement for anatomical fidelity or evaluate opposition with orientation-aware metrics. Likewise, soft tissue structures like the thenar eminence (TE) and first web space (FWS) are frequently omitted or included only cosmetically, limiting ergonomic grip quality. Here, we introduce a design and evaluation workflow for anatomical thumb replication in robotic hands. Using a compact parameter space, a MATLAB framework systematically optimizes TMC joint geometry for human-like motion. To evaluate grasping performance, we propose the Kapandji-plus test, which augments standard reachability assessments with orientation scoring across 17 functional poses. Additionally, we developed 21 novel passive TE and FWS prototypes with varied compliance and geometry, tested using task-oriented quantitative and qualitative metrics. Results show that two optimized joint configurations significantly improve thumb orientation accuracy and Kapandji-plus scores over a baseline configuration. The best-performing TE and FWS designs enhance power and precision grips, with a kirigami-inspired FWS demonstrating repeatable lateral support in manipulation tasks. These components advance anthropomorphic hand design by combining anatomically informed mechanics with task-oriented validation.

Keywords: Representation Learning, Humanoid and Bipedal Locomotion, Legged Robots
Abstract: Reinforcement learning (RL), driven by data-driven methods, has become an effective solution for robot leg motion control problems. However, the mainstream RL methods for bipedal robot terrain traversal, such as teacher-student policy knowledge distillation, suffer from long training times, which limit development efficiency. To address this issue, this paper proposes BarlowWalk, an improved Proximal Policy Optimization (PPO) method integrated with self-supervised representation learning. This method employs the Barlow Twins algorithm to construct a decoupled latent space, mapping historical observation sequences into low-dimensional representations and implementing self-supervision. Meanwhile, the actor requires only proprioceptive information to achieve self-supervised learning over continuous time steps, significantly reducing the dependence on external terrain perception. Simulation experiments demonstrate that this method has significant advantages in complex terrain scenarios. To enhance the credibility of the evaluation, this study compares BarlowWalk with advanced algorithms through comparative tests, and the experimental results verify the effectiveness of the proposed method.

Keywords: Humanoid and Bipedal Locomotion, Hydraulic/Pneumatic Actuators, Actuation and Joint Mechanisms
Abstract: Toward realization of lightweight and durable legged robots, a new hydraulic leg was designed, fabricated, and tested. The four-joint leg is 0.9 m in height, and 9.5 kg in weight. The featured design lies in bundling multiple cylinders, valves and sensors into one unit, then actuating some joints proximally via mechanical linkages. The integration allows for internal routing of hydraulic pipelines and structural protrusions, enabling a sensor-less configuration beyond the knee joint, significantly decreasing the risk of damage during impacts or falls. The each hydraulic servo actuator incorporates pressure feedback to realize joint torque control. The proposed design concept and the realized system were validated through preliminary experiments using a single-leg robot. The prototype robot demonstrated soft landing from a height of 0.3 m while carrying a 30 kg payload without sustaining damage, and achieved a vertical jump of 0.27 m without payload. The results are also explained in the supplementary video.

Keywords: Multi-Contact Whole-Body Motion Planning and Control, Manipulation Planning, Dual Arm Manipulation
Abstract: Daily tasks require us to use our whole body to manipulate objects, for instance when our hands are unavailable. We consider the issue of providing humanoid robots with the ability to autonomously perform similar whole-body manipulation tasks. In this context, the infinite possibilities for where and how contact can occur on the robot and object surfaces hinder the scalability of existing planning methods, which predominantly rely on discrete sampling. Given the continuous nature of contact surfaces, gradient-based optimization offers a more suitable approach for finding solutions. However, a key remaining challenge is the lack of an efficient representation of robot surfaces. In this work, we propose (i) a representation of robot and object surfaces that enables closed-form computation of proximity points, and (ii) a cost design that effectively guides whole-body manipulation planning. Our experiments demonstrate that the proposed framework can solve problems unaddressed by existing methods, and achieves a 77% improvement in planning time over the state of the art. We also validate the suitability of our approach on real hardware through the whole-body manipulation of boxes by a humanoid robot.

Keywords: Body Balancing, Reinforcement Learning, Humanoid Robot Systems
Abstract: Maintaining balance under disturbances is a key challenge for humanoid robots due to their instability and complexity. Traditional controllers, relying on simplified models, struggle to adapt in dynamic situations. We present a deep reinforcement learning approach for push recovery on the Talos humanoid, using a reward shaped by capture point and divergent component of motion theories. The resulting controller exhibits rapid convergence during training and strong generalization to different push scenarios and simulations. These results highlight the importance of physically informed reward design for robust, adaptable balance control in humanoid robots.

Keywords: Grasping, Multifingered Hands, Force and Tactile Sensing
Abstract: The control of interaction forces is crucial for successful grasping and manipulation. Especially in multifingered hands, the hand posture needs to be coordinated in order to achieve a contact force pattern that ensures grasp stability and fulfills any task-related requirements. We propose a novel hand model that describes grasp contact locations in correlation to the hand posture. Based on the concept of grasp opposition types we identify the main areas of force interaction between the hand and the object from human demonstrations during a manipulation task. This allows us to identify and categorize different force opposition types and leverage these for the control of a robotic hand.

Keywords: Physical Human-Robot Interaction, Touch in HRI, Multi-Modal Perception for HRI
Abstract: Robots that can physically interact with humans in a safe, comfortable, and intuitive manner can help in a variety of settings. In this paper we propose a perception-based interaction adaptation framework. One main component of this framework is a multi-modal perception model which is grounded on the existing literature and is intended to provide a quantitative estimation of the human state- defined as the perceptions of the physical interaction- by using human, robot, and context information. The estimated human state is fed to a context-aware behavior adaptation framework which recommends robot behaviors to improve human state using a learned behavior cost model and an optimization formulation. We show the potential of such a human state estimation model by evaluating a reduced model through in-person user studies.

Keywords: Dual Arm Manipulation, Compliant Assembly, Humanoid Robot Systems
Abstract: Delicate snap-fit assemblies, such as those in precision fits like inserting a lens into an eyewear frame or in electronics, demand timely engagement detection and rapid force attenuation to prevent overshoot-induced component damage or assembly failure. In this work, we introduce a bimanual manipulation framework that integrates both capabilities into a unified skill for snap-fit assembly. The system relies solely on joint-level proprioception: a learned model detects engagement from joint-velocity transients and subsequently triggers a task-aware stiffness modulation along the insertion axis. The bimanual policy is structured around a coupled dynamical system (DS) that coordinates synchronized transport motion with selective decoupling during insertion. We evaluate the framework across varied geometries and robot platforms, demonstrating its applicability to real-world snap-fit tasks.

Keywords: Soft Robot Materials and Design, Soft Sensors and Actuators, Modeling, Control, and Learning for Soft Robots
Abstract: The muscle-tendon complex (MTC) integrates contractile actuation and proprioceptive sensing of the muscle length and the muscle tension, which enables the feedback control of muscle activations in biological systems. Artificial muscles, often utilizing polymeric materials with intricate material behaviors, need to incorporate such proprioceptive functionalities for reliable and adaptable feedback control. Here, we present an MTC-inspired liquid crystal elastomer (LCE) artificial muscles that can simultaneously actuate through Joule heating and sense the mechanical states of the LCE with embedded liquid metal (LM) channels. This is enabled by a multi-material design consisting of isotropic LCE (iso-LCE) and nematic LCE (nem-LCE) with different thermomechanical properties suited to specific functions of the LM –– joule heating and sensing of force and stretch. All sensing and actuating components are integrated into a single specimen, creating a compact, compliant, and proprioceptive artificial muscle. Furthermore, the LCE actuators are configured to form an antagonistic pair that closely resembles the structure of biological muscles, improving the controllability of the LCE actuators. These MTC-inspired LCE artificial muscles are used to perform closed-loop feedback control of a robotic finger and a gripper system, demonstrating the potential applications of the LCE actuators in enhanced robotic control system.

Keywords: Sensorimotor Learning, Reinforcement Learning, Representation Learning
Abstract: An important element of learning contact-rich manipulation for robots is leveraging the synergy between heterogeneous sensor modalities such as vision, force, and proprioception while adapting to sensory perturbations and dynamic changes. In such multisensory settings, Reinforcement Learning (RL) faces caveats arising from varying sensory feature distributions and their changing importance depending on the task phase. In this work, we propose MultiSensory Dynamic Pretraining (MSDP), a novel framework for learning multisensory representations tailored for task-oriented policy learning via masked autoencoding coupled with self-supervised forward dynamics objectives to shape features from multiple different sensors.

Keywords: Multi-Robot Systems, Reinforcement Learning, Human-Robot Teaming
Abstract: Humanoid robots are increasingly relying on re- inforcement learning by building reward models aligned to humans, such as training language models to follow human instructions. However, multi-agent reinforcement learning sys- tems such as robot teaming suffers large performance loss due to reward model anomaly, and their low observability makes anomaly detection challenging. This paper investigates the im- pact of poisoning attacks that exploit shared reward structures in multi-agent reinforcement learning, luring agents into reward traps. Specifically, we present a poisoning attack tailored for deep reinforcement learning in multi-agent setup, and evaluate its vulnerability on two representative reinforcement learning algorithms: PPO and SAC. Results demonstrate performance degradation of 18.7% (PPO) and 20.9 % (SAC). While SAC showed a marginal decline in performance compared to PPO, PPO experienced a severe reward collapse under attack. This suggests that PPO is vulnerable to poisoning attacks, especially in multi-agent environments. These findings call for robust defense mechanisms against reward-based attacks in multi- agent reinforcement learning systems.

Keywords: Grasping, Dexterous Manipulation, Performance Evaluation and Benchmarking
Abstract: The goals of COMPARE Ecosystem are to (1) create a greater cohesion between open-source products by generating community-driven standards for components of software pipelines that increase modularity and enable implementation that allows for greater performance quantification, (2) unite the existing community of users and developers to build upon and integrate existing open-source products as well as improve and evolve the future of robot manipulation research, and (3) facilitate sufficient commonality in hardware and software that allows for quantitative evaluation of research and eases the implementation of the complex robot manipulation pipeline. The ecosystem will produce standards and guidelines for developing open-source products, conducting benchmarking evaluations, and reporting performance results. Test and evaluation procedures will be developed to validate the open-source products and their compatibility with others in the software pipeline, including reviews of their adherence to the developed standards and guidelines, performance assessments of their functionality, and usability analyses when attempting to integrate the products. With an initial use case for picking in clutter, COMPARE is structured around open-source software components (grasp planning, motion planning, perception, learning, and simulation), open-source hardware designs (arms, end-effectors, tactile sensors, and humanoids), and benchmarking assets (datasets, objects and artifacts, tools and testbeds, protocols, and leaderboards). All materials are hosted on Robot-Manipulation.org, a landing page for the ecosystem featuring actively maintained repositories of open-source products, benchmarking assets, and events like workshops and competitions. We aim to continue growing the COMPARE Ecosystem and to expand to include more applications including dexterous manipulation, contact-rich manipulation, soft robotics, humanoids, and more.

Keywords: Bimanual Manipulation, Deep Learning in Grasping and Manipulation, Deep Learning for Visual Perception
Abstract: For the task of hanging clothes, learning how to insert a hanger into a garment is a crucial step, but has rarely been explored in robotics. In this work, we address the problem of inserting a hanger into various unseen garments that are initially laid flat on a table. This task is challenging due to its long-horizon nature, the high degrees of freedom of the garments and the lack of data. To simplify the learning process, we first propose breaking the task into several subtasks. Then, we formulate each subtask as a policy learning problem and propose a low-dimensional action parameterization. To overcome the challenge of limited data, we build our own simulator and create 144 synthetic clothing assets to effectively collect high-quality training data. Our approach uses single-view depth images and object masks as input, which mitigates the Sim2Real appearance gap and achieves high generalization capabilities for new garments. Extensive experiments in both simulation and reality validate our proposed method. By training on various garments in the simulator, our method achieves a 75% success rate with 8 different unseen garments in the real world.

Keywords: Humanoid Robot Systems, Compliant Joints and Mechanisms, Dexterous Manipulation
Abstract: Dexterity and strength are essential for performing a variety of tasks in the unstructured environments of household services and craftsmanship. For these tasks, we have developed neoDavid, a robust humanoid robot with dexterous manipulation skills. neoDavid has joints with variable stiffness actuators (VSA) which have mechanically adjustable elasticity in the drive train, a continuum elastic neck, and a gravitationally compensated torso with overload couplings. We present our modular approach in the development starting from system architecture over mechatronic components, communication, and control to higher-level software. We demonstrate how this modularity enables scalable enhancements, allowing us to evolve the system from a single arm and hand into a complete humanoid robot. Additionally, we highlight advancements in perception, manipulation, and motion planning, along with the implementation of offline task planning utilizing capability maps. The versatile character in terms of dexterity and robustness is demonstrated in challenging applications, e.g., handling a drill hammer, fine manipulation of a pipette, and emptying a dishwasher.

Keywords: Telerobotics and Teleoperation, Humanoid Robot Systems, Human-Robot Teaming
Abstract: Achieving seamless synchronization between user and robot motion in teleoperation, particularly during high-speed tasks, remains a significant challenge. In this work, we propose a novel approach for transferring stepping motions from the user to the robot in real-time. Instead of directly replicating user foot poses, we retarget user steps to robot footstep locations, allowing the robot to utilize its own dynamics for locomotion, ensuring better balance and stability. Our method anticipates user footsteps to minimize delays between when the user initiates and completes a step and when the robot does it. The step estimates are continuously adapted to converge with the measured user references. Additionally, the system autonomously adjusts the robot’s steps to account for its surrounding terrain, overcoming challenges posed by environmental mismatches between the user's flat-ground setup and the robot's uneven terrain. Experimental results on our humanoid robot demonstrate the effectiveness of the proposed system.

Keywords: Human and Humanoid Motion Analysis and Synthesis, Humanoid Robot Systems
Abstract: Scalable learning of humanoid robots is crucial for their deployment in real-world applications. While traditional approaches primarily rely on reinforcement learning or teleoperation to achieve whole-body control, they are often limited by the diversity of simulated environments and the high costs of demonstration collection. In contrast, human videos are ubiquitous and present an untapped source of semantic and motion information that could significantly enhance the generalization capabilities of humanoid robots. This paper introduces Humanoid-X, a large-scale dataset of over 20 million humanoid robot poses with corresponding text-based motion descriptions, designed to leverage this abundant data. Humanoid-X is curated through a comprehensive pipeline: data mining from the Internet, video caption generation, motion retargeting of humans to humanoid robots, and policy learning for real-world deployment. With Humanoid-X, we further train a large humanoid model, UH-1, which takes text instructions as input and outputs corresponding actions to control a humanoid robot. Extensive simulated and real-world experiments validate that our scalable training approach leads to superior generalization in text-based humanoid control, marking a significant step toward adaptable, real-world-ready humanoid robots.

Keywords: Imitation Learning, AI-Based Methods, Reinforcement Learning
Abstract: Imitation learning is a promising approach for enabling generalist capabilities in humanoid robots, but its scaling is fundamentally constrained by the scarcity of high-quality expert demonstrations. This limitation can be mitigated by leveraging suboptimal, open-ended play data, often easier to collect and offering greater diversity. This work builds upon recent advances in generative modeling, specifically Flow Matching, an alternative to Diffusion models. We introduce a method for estimating the extremum of the learned distribution by leveraging the unique properties of Flow Matching, namely, deterministic transport and support for arbitrary source distributions. We apply this method to develop several goal-conditioned imitation and reinforcement learning algorithms based on Flow Matching, where policies are conditioned on both current and goal observations. We explore and compare different architectural configurations by combining core components, such as critic, planner, actor, or world model, in various ways. We evaluated our agents on the OGBench benchmark and analyzed how different demonstration behaviors during data collection affect performance in a 2D non-prehensile pushing task. Furthermore, we validated our approach on real hardware by deploying it on the Talos humanoid robot to perform complex manipulation tasks based on high-dimensional image observations, featuring a sequence of pick-and-place and articulated object manipulation in a realistic kitchen environment. Experimental videos and code are available at: https://hucebot.github.io/extremum_flow_matching_website/

Keywords: Behavior-Based Systems, Human-Robot Collaboration
Abstract: Autonomous elevator navigation represents a critical challenge for service robots operating in multi-floor environments. This paper presents a novel framework that integrates autonomous elevator operation with human assistance, enabling robots to navigate elevators in diverse scenarios with varying human presence. Our approach determines whether autonomous operation is feasible based on real-time environmental constraints and reactively switches to seeking human help when necessary. We demonstrate how combining different navigation cost metrics allows the robot to navigate safely among humans and reliably detect door states based on LiDAR data, even with humans entering or exiting the elevator. We validate our system through comprehensive testing in both simulated human-robot interaction scenarios and real robot experiments using the humanoid household robot ARMAR-7. Results show significantly improved success rates across diverse elevator situations compared to pure autonomous or help-seeking baselines.

Keywords: Hardware-Software Integration in Robotics, Humanoid and Bipedal Locomotion, Humanoid Robot Systems
Abstract: Various bipedal robots have been developed to date, and in recent years, there has been a growing trend toward releasing these robots as open-source platforms. This shift is fostering an environment in which anyone can freely develop bipedal robots and share their knowledge, rather than relying solely on commercial products. However, most existing open-source bipedal robots are designed to be fabricated using 3D printers, which limits their scalability in size and often results in fragile structures. On the other hand, some metal-based bipedal robots have been developed, but they typically involve a large number of components, making assembly difficult, and in some cases, the parts themselves are not readily available through e-commerce platforms. To address these issues, we developed MEVITA, an open-source bipedal robot that can be built entirely from components available via e-commerce. Aiming for the minimal viable configuration for a bipedal robot, we utilized sheet metal welding to integrate complex geometries into single parts, thereby significantly reducing the number of components and enabling easy assembly for anyone. Through reinforcement learning in simulation and Sim-to-Real transfer, we demonstrated robust walking behaviors across various environments, confirming the effectiveness of our approach. All hardware, software, and training environments can be obtained from https://github.com/haraduka/mevita.

Keywords: Physical Human-Robot Interaction
Abstract: We present a hierarchical policy‐learning framework that enables a legged humanoid to cooperatively carry extended loads with a human partner using only haptic cues for intent inference. At the upper tier, a lightweight behavior‐cloning network consumes six‐axis force/torque streams from dual wrist‐mounted sensors and outputs whole‐body planar velocity commands that capture the leader’s applied forces. At the lower tier, a deep‐reinforcement‐learning policy, trained under randomized payloads (0–3 kg) and friction conditions in Isaac Gym and validated in MuJoCo and on a real Unitree G1—maps these high‐level twists to stable, under‐load joint trajectories. By decoupling intent interpretation (force → velocity) from legged locomotion (velocity → joints), our method combines intuitive responsiveness to human inputs with robust, load‐adaptive walking. We collect training data without motion‐capture or markers—only synchronized RGB video and F/T readings—employing SAM2 and WHAM to extract 3D human pose and velocity. In real‐world trials, our humanoid achieves cooperative carry‐and‐move performance (completion time, trajectory deviation, velocity synchrony, and follower‐force) on par with a blindfolded human‐follower baseline. This work is the first to demonstrate learned haptic guidance fused with full‐body legged control for fluid human–humanoid co–manipulation. Code and videos are on our href{https://h2compact.github.io/h2compact/}{website}.

Keywords: Legged Robots, Mechanism Design, Methods and Tools for Robot System Design
Abstract: A monoped’s jump height and energy consumption depend on both, its mechanical design and control strategy. Existing co-design frameworks typically optimize for either maximum height or minimum energy, neglecting their trade-off. They also often omit gearbox parameter optimization and use oversimplified actuator mass models, producing designs difficult to replicate in practice. In this work, we introduce a novel three-stage co-design optimization framework that jointly maximizes jump height while minimizing mechanical energy consumption of a monoped. The proposed method explicitly incorporates realistic actuator mass models and optimizes mechanical design (including gearbox) and control parameters within a unified framework. The resulting design outputs are then used to automatically generate a parameterized CAD model suitable for direct fabrication, significantly reducing manual design iterations. Our experimental evaluations demonstrate a 50% reduction in mechanical energy consumption compared to the baseline design, while achieving a jump height of 0.8m.

Keywords: Mobile Manipulation, Social HRI, Safety in HRI
Abstract: We present Lucio, a competition-validated, open-hardware mobile manipulator designed to execute complex domestic tasks with robust autonomy and full on-device reasoning. Lucio combines a holonomic omnidirectional base, a 7-DoF Kinova Gen3 arm, and a Robotiq gripper with a modular ROS-based software framework that scales across wheeled, legged, and humanoid platforms. Its perception pipeline leverages YOLOv11, point-cloud filtering, and colored heightmap grasp sampling to perform reliable object detection and manipulation, even in cluttered environments.

At the core of Lucio’s high-level control is the Robotic Decision-Making Model (RDMM)—a 4-bit quantized large language model fine-tuned on 27,000 task plans and 1,300 multimodal instruction-action pairs derived from RoboCup@Home scenarios. Running entirely on an embedded Jetson Orin, RDMM enables low-latency natural language task execution without cloud reliance. Lucio was validated in two full RoboCup@Home seasons, consistently completing complex tasks such as Storing Groceries, Clean the Table, and General Purpose Service Robot with zero manual resets.

We release Lucio’s full hardware design, software stack, and simulation environments to support reproducible research in service robotics, task planning, and on-device LLM integration.

Keywords: Humanoid and Bipedal Locomotion, Reinforcement Learning, Legged Robots
Abstract: This paper presents a reinforcement learning framework for navigation-free point-to-point walking in bipedal robots. Unlike traditional velocity-command-based approaches, we introduce displacement-based commands that align more naturally with the discrete stepping nature of legged locomotion. Our method enables smooth transitions between standing and walking, accurate control over the number of steps, and support for both discrete and continuous walking trajectories. We design a phase-encoded command format and train the policy using a history of proprioceptive states and well-crafted reward functions. The policy is trained with PPO in Isaac Lab and deployed on a full-size humanoid robot (1.8 meters tall, 80 kg, 6 DoF per leg). In simulation, the controller achieves an average tracking error of less than 0.11 meters in position and 0.036 radians in heading. On hardware, the errors increase to approximately 0.3 meters and 0.1 radians due to sim-to-real discrepancies. The proposed framework demonstrates robust and efficient point-to-point locomotion without the need for high-level navigation modules.

Keywords: Humanoid and Bipedal Locomotion, Reinforcement Learning, Machine Learning for Robot Control
Abstract: Achieving stable and robust perceptive locomotion for bipedal robots in unstructured outdoor environments remains a critical challenge due to complex terrain geometry and susceptibility to external disturbances. In this work, we propose a novel reward design inspired by the Linear Inverted Pendulum Model (LIPM) to enable perceptive and stable locomotion in the wild. The LIPM provides theoretical guidance for dynamic balance by regulating the center of mass (CoM) height and the torso orientation. These are key factors for terrain-aware locomotion, as they help ensure a stable viewpoint for the robot's camera. Building on this insight, we design a reward function that promotes balance and dynamic stability while encouraging accurate CoM trajectory tracking. To adaptively trade off between velocity tracking and stability, we leverage the Reward Fusion Module (RFM) approach that prioritizes stability when needed. A double-critic architecture is adopted to separately evaluate stability and locomotion objectives, improving training efficiency and robustness. We validate our approach through extensive experiments on a bipedal robot in both simulation and real-world outdoor environments. The results demonstrate superior terrain adaptability, disturbance rejection, and consistent performance across a wide range of speeds and perceptual conditions.

Keywords: Data Sets for Robot Learning, Big Data in Robotics and Automation, Social HRI
Abstract: Robotic decision-making with large language models (LLMs) is increasingly constrained by the scarcity of datasets that are both linguistically expressive and grounded in symbolic robotic actions. We introduce the Dynamic RDMM Dataset, a controllable, text-to-text dataset generator that maps natural-language instructions to structured action programs across 23 real-world household tasks.

RDMM is built using a two-stage generation process. First, hierarchical template expansion recursively constructs multi-step task descriptions from nested logic templates, enabling variable instruction complexity and compositional diversity. Second, constraint-aware content generation fills these templates using curated embeddings (verbs, rooms, objects, names) while enforcing semantic and physical validity. This process produces 1,800 expert-verified instruction–action pairs and can be scaled to over 100,000 diverse, syntactically and semantically valid examples.

Crucially, RDMM supports task rebalancing and adjustable instruction difficulty, allowing researchers to generate lightweight commands (e.g., "go to the kitchen") or compound sequences (e.g., "deliver an item to Kai and follow him until he sits"). This makes RDMM a practical tool for curriculum learning, ablation studies, and controlled benchmarking.

We validate RDMM by fine-tuning LLaMA-3-8B, Mistral-7B, and Qwen-0.5 models, achieving over 94% accuracy and outperforming ChatGPT-4o baselines. In real-world deployment at RoboCup@Home, RDMM-trained models successfully executed unseen multi-step commands in noisy, natural environments. RDMM is released with templates, code, and generation tools to support reproducible, instruction-grounded robot learning.

Keywords: Multi-Modal Perception for HRI, Touch in HRI, Physical Human-Robot Interaction
Abstract: This paper introduces a humanoid robotic bust designed for natural, real-time interaction using multiple sensory inputs—vision, sound, and touch. The system brings together dedicated hardware and control architectures to respond to human presence and contact in a socially intuitive way. A custom-built tactile sensor embedded in the face detects both the location and strength of touch, while an eye-mounted camera and deep learning model guide gaze-based responses to visual cues. Directional sounds trigger head movements through auditory-based servoing, allowing the robot to react to its surroundings dynamically. Each sensory input is mapped to specific motor behaviors, enabling fluid, lifelike responses driven by a three-degree-of-freedom actuation module. Demonstration scenarios highlight the system’s ability to perform socially appropriate and responsive behavior. These results suggest a strong foundation for future work involving multi-sensory fusion, speech interaction, and more expressive motion control.

Keywords: Legged Robots, Mechanism Design, Humanoid Robot Systems
Abstract: We propose a novel leg structure with concentrated liquid-cooling actuators for long-term operation of humanoid robots. Many humanoid robots recently developed employ quasi-direct drive actuators to achieve high dynamic capabilities. This design choice requires high power consumption in return, leading to significant heat generation. Liquid-cooling systems are effective in efficiently removing heat from actuators, making them beneficial for enhancing/keeping the performance of humanoid robots. However, the systems come with unignorable drawbacks, including increased system complexity due to tube routing and the potential for catastrophic damage from liquid leakage, especially when mounting actuators on the ankle as in conventional leg design. To mitigate this, we propose a structure with ``concentrated'' liquid-cooling actuators, placing all actuators above the knee and cooling them collectively. This structure reduces both leakage risk and component count while retaining the benefits of liquid-cooling. Experiments for long-term operation and impact resistance were conducted both in simulations using a humanoid model equipped with the proposed structure, and in the physical world using a leg with the proposed structure. The results demonstrated that the proposed structure enables operation at lower actuator temperatures and allows the robot to perform impact-intensive tasks, such as ball kicking, without any issues.

Keywords: Telerobotics and Teleoperation, Haptics and Haptic Interfaces, Multifingered Hands
Abstract: This paper presents GD-Prox, a Haptic-Glove Dual-Proxy framework for remote dexterous robot hand telemanipulation. The proposed method extends the Dual-Proxy approach to glove–hand systems, addressing inherent actuation constraints and enabling stable bilateral interaction over low-bandwidth networks without direct task-space force transmission. Independent local controllers on the glove and robot-hand sides maintain responsive behavior under significant communication delays and reduced update rates. Experimental validation using intermittent press–release and sustained pinch–grasp tasks confirms the effectiveness of the approach in maintaining stable bilateral coordination under latency. The framework establishes a robust foundation for teleoperating robot hands with limited degrees of freedom, paving the way for future extensions to inter-finger coordination and dynamic object manipulation.

Keywords: Humanoid and Bipedal Locomotion, Legged Robots, Underactuated Robots
Abstract: As humanoid robots enter real-world environments, ensuring robust locomotion across diverse environments is crucial. This paper presents a computationally efficient hierarchical control framework for humanoid robot locomotion based on reduced-order models—enabling versatile step planning and incorporating arm and torso dynamics to better stabilize the walking. At the high level, we use the step-to-step dynamics of the ALIP model to simultaneously optimize over step periods, step lengths, and ankle torques via nonlinear MPC. The ALIP trajectories are used as references to a linear MPC framework that extends the standard SRB-MPC to also include simplified arm and torso dynamics. We validate the performance of our approach through simulation and hardware experiments on the Unitree G1 humanoid robot. In the proposed framework the high-level step planner runs at 40 Hz and the mid-level MPC at 500 Hz using the onboard mini-PC. Adaptive step timing increased the push recovery success rate by 36%, and the upper body control improved the yaw disturbance rejection. We also demonstrate robust locomotion across diverse indoor and outdoor terrains, including grass, stone pavement, and uneven gym mats.

Keywords: Humanoid and Bipedal Locomotion, Legged Robots
Abstract: Humanoids operating in real-world workspaces must frequently execute task-driven, short-range movements to SE(2) target poses. To be practical, these transitions must be fast, robust, and energy efficient. While learning-based locomotion has made significant progress, most existing methods optimize for velocity-tracking rather than direct pose reaching, resulting in inefficient, marching-style behavior when applied to short-range tasks. In this work, we develop a reinforcement learning approach that directly optimizes humanoid locomotion for SE(2) targets. Central to this approach is a new constellation-based reward function that encourages natural and efficient target-oriented movement. To evaluate performance, we introduce a benchmarking framework that measures energy consumption, time-to-target, and footstep count on a distribution of SE(2) goals. Our results show that the proposed approach consistently outperforms standard methods and enables successful transfer from simulation to hardware, highlighting the importance of targeted reward design for practical short-range humanoid locomotion.

Keywords: Bioinspired Robot Learning, Humanoid and Bipedal Locomotion, Reinforcement Learning
Abstract: We present a unified gait-conditioned reinforcement learning (RL) framework that enables humanoid robots to perform standing, walking, running, and smooth transitions within a single recurrent policy. A compact reward routing mechanism dynamically activates gait-specific objectives based on a one-hot gait ID, mitigating reward interference and supporting stable multi-gait learning. Human-inspired reward terms promote biomechanically natural motions, such as straight-knee stance and coordinated arm-leg swing, without requiring motion capture data. A structured curriculum progressively introduces gait complexity and expands command space over multiple phases. In simulation, the policy successfully achieves robust standing, walking, running, and gait transitions. On the real Unitree G1 humanoid, we validate standing, walking, and walk-to-stand transitions, demonstrating stable and coordinated locomotion. This work provides a scalable, reference-free solution toward versatile and naturalistic humanoid control across diverse modes and environments.

Keywords: Humanoid Robot Systems, Passive Walking, Reinforcement Learning
Abstract: Bipedal walking is a fundamental mode of locomotion in nature and has long inspired the development of humanoid robots. Passive dynamic walking stands out among bipedal locomotion strategies for its energy efficiency and human-like gait, but suffers from low stability and sensitivity to terrain. In this paper, we propose a novel method for realizing passive walking on a physical robot by combining a kneed walker model, a variant of passive dynamic walking, with reinforcement learning. Specifically, we develop a hierarchical control framework in which high-level behaviors are planned using a simplified passive walking model, while low-level control is handled by a reinforcement learning policy trained to bridge the gap between the idealized model and real-world dynamics. The method is validated in a Sim2Sim using two different robot platforms in the MuJoCo simulation environment. The proposed approach is compared to other approaches, demonstrating that it significantly reduces energy consumption - by approximately 16% and 31% in cost of transport (CoT) for two types of robots. Additionally, it generates more natural and stable motions, as shown by lower ground reaction forces.

Keywords: Humanoid and Bipedal Locomotion, Machine Learning for Robot Control, Reinforcement Learning
Abstract: Sim-to-real reinforcement learning (RL) for humanoid robots with high-gear ratio actuators remains challenging due to complex actuator dynamics and the absence of torque sensors. To address this, we propose a novel RL framework leveraging foot-mounted inertial measurement units (IMUs). Instead of pursuing detailed actuator modeling and system identification, we utilize foot-mounted IMU measurements to enhance rapid stabilization capabilities over challenging terrains. Additionally, we propose symmetric data augmentation dedicated to the proposed observation space and random network distillation to enhance bipedal locomotion learning over rough terrain. We validate our approach through hardware experiments on a miniature-sized humanoid EVAL-03 over a variety of environments. The experimental results demonstrate that our method improves rapid stabilization capabilities over non-rigid surfaces and sudden environmental transitions.

Keywords: Service Robotics, Mobile Manipulation
Abstract: In this paper, we present a whole-body impedance controller for mobile service robots equipped with non-holonomic differential-drive bases and dual arms. We first derive the constrained whole-body dynamics of the robot and then reduce the model to a lower-dimensional space to incorporate the non-holonomic constraints of the mobile base. Based on this reduced model, impedance control is implemented in the task space. Secondary control objectives are realized in the robot's null space, including PD control for the robot's home configuration, joint limit avoidance, singularity avoidance, and self-collision prevention. The controller's performance is validated through both simulation and experimental studies on the GARMI service robot.

Keywords: Domestic Robotics, Perception for Grasping and Manipulation, Legged Robots
Abstract: Despite increasing research efforts on household robotics, robots intended for deployment in domestic settings still struggle with more complex tasks such as interacting with functional elements like drawers or light switches, largely due to limited task-specific understanding and interaction capabilities. These tasks require not only detection and pose estimation but also an understanding of the affordances these elements provide. To address these challenges and enhance robotic scene understanding, we introduce SpotLight: A comprehensive framework for robotic interaction with functional elements, specifically light switches. Furthermore, this framework enables robots to improve their environmental understanding through interaction. Leveraging VLM-based affordance prediction to estimate motion primitives for light switch interaction, we achieve up to 84% operation success in real-world experiments. We further introduce a specialized dataset containing 715 images as well as a custom detection model for light switch detection. We demonstrate how the framework can facilitate robot learning through physical interaction by having the robot explore the environment and discover previously unknown relationships in a scene graph representation. Lastly, we propose an extension to the framework to accommodate other functional interactions such as swing doors, showcasing its flexibility.

Keywords: Deep Learning in Grasping and Manipulation, Perception for Grasping and Manipulation, Deep Learning for Visual Perception
Abstract: Enabling robots to grasp objects specified by natural language is crucial for effective human–robot interaction, yet remains a significant challenge. Existing approaches often struggle with open–form language expressions and assume unambiguous target objects without duplicates. Moreover, they frequently depend on costly, dense pixel–wise annotations for both object grounding and grasp configuration. We present Attribute–based Object Grounding and Robotic Grasping (OGRG), a novel model that interprets open–form language expressions and performs spatial reasoning to ground targets and predict planar grasp poses, even in scenes with duplicated objects. We investigate OGRG in two settings: (1) Referring Grasp Synthesis (RGS) under pixel–wise full supervision, and (2) Referring Grasp Affordance (RGA) using weakly supervised learning that requires only single–pixel grasp annotations. Key contributions include a bi–directional vision–language fusion module and the integration of depth information for improved geometric reasoning, enhancing both grounding and grasping performance. Experiment results demonstrate superior performance over strong baselines in tabletop scenes with varied spatial language instructions. For RGS, OGRG operates at 17.59 FPS on a single NVIDIA RTX 2080 Ti GPU, enabling potential use in closed–loop or multi–object sequential grasping, while delivering higher grounding and grasp–prediction accuracy than all baseline methods. Under the weakly supervised RGA setting, the model likewise surpasses baseline grasp–success rates in both simulation and real–robot trials, underscoring the effectiveness of its spatial–reasoning design.

Keywords: Big Data in Robotics and Automation, Methods and Tools for Robot System Design, Learning from Experience
Abstract: This paper introduces RT-cache, a novel trajectory-memory pipeline that accelerates real-world robot inference by leveraging big-data retrieval and learning from experience. While modern Vision-Language-Action (VLA) models can handle diverse robotic tasks, they often incur high per-step inference costs, resulting in significant latency, sometimes minutes per task. In contrast, RT-cache stores a large-scale Memory of previously successful robot trajectories and retrieves relevant multi-step motion snippets, drastically reducing inference overhead. By integrating a Memory Builder with a Trajectory Retrieval, we develop an efficient retrieval process that remains tractable even for extremely large datasets. RT-cache flexibly accumulates real-world experiences and replays them whenever the current scene matches past states, adapting quickly to new or unseen environments with only a few additional samples. Experiments on the Open-X Embodiment Dataset and other real-world data demonstrate that RT-cache completes tasks both faster and more successfully than a baseline lacking retrieval, suggesting a practical, data-driven solution for real-time manipulation.

Keywords: Bimanual Manipulation, Dual Arm Manipulation
Abstract: In this work, the problem of robotic shelf replenishment is being studied. Such tasks involve a variety of object types and geometries, which should be picked from boxes where they are tightly packed and placed on shelves in tight formations with the appropriate orientation. We consider a bimanual robotic set up with a parallel finger gripper and a 3D-printed end-effector and propose to combine simple grasping with a set of non-prehensile manipulations to achieve such a replenishment task. This work reports on the details of the implementation of the proposed strategy and on the initial investigation of the feasibility and effectiveness of the proposed solution for a representative set of super-market products, demonstrated in a number of experiments in the lab.

Keywords: Human and Humanoid Motion Analysis and Synthesis, Modeling and Simulating Humans, Humanoid and Bipedal Locomotion
Abstract: This study identifies a robust parameter for quantifying instability in general biped systems by comparing mechanical stability descriptors in humans and humanoid biped robots during locomotion. Three key parameters were investigated: the distance between the center of mass to the minimal moment axis (dCoM-MMA), the margin of stability (MoS), and whole-body angular momentum (WBAM). Using both robot experiments and human walking trials under normal and perturbed conditions, we analyzed how these descriptors behave under varying levels of instability. Our findings demonstrate that dCoM-MMA is a sensitive and robust metric for detecting instability across both humans and biped robots, independent on gait variability, perturbation levels, and methods. Comparative analyses show that dCoM-MMA is predictive of different levels of instability, highlighting its potential use in unified stability analysis across both fields. This research provides insights that could be beneficial to the design of exoskeletons, fall monitoring systems, and gait-assistive devices, thereby addressing the needs of aging populations.

Keywords: Humanoid and Bipedal Locomotion, Legged Robots, Neural and Fuzzy Control
Abstract: This study proposes a robust and generalised contact detection framework for bipedal robots based on foot IMUs. The framework employs a hierarchical fuzzy inference approach to identify contact with the ground and sliding. The proposed approach integrates a general type-2 fuzzy clustering mean for contact detection with an interval type-2 fuzzy logic system for slide detection. This integration enables the model to handle multidimensional uncertainty and generalise across a wide range of terrain and gait conditions. A further advantage of the foot-mounted IMU approach is that it eliminates the need for external contact sensors or force transducers. In comparison to black-box deep learning approaches, this approach offers enhanced interpretability and generalization capabilities. The efficacy and versatility of the framework are demonstrated by simulation experiments, which show an average contact detection accuracy exceeding 95% in both structured and unstructured terrain scenarios without external contact force sensors or large datasets for the training required.

Keywords: Prosthetics and Exoskeletons, Rehabilitation Robotics, Machine Learning for Robot Control
Abstract: This work introduces a controller for an upper-limb rehabilitative exoskeleton based on reservoir computing (RC). The controller decodes the motor intention of the user by observing the electromyographic (EMG) activity of four upper-limb muscles and end effector (EE) kinematics and then assists the movements of upper-limb during the execution of planar reaching tasks. After tuning the hyperparameters of the RC, the controller was tested by three healthy participants wearing a shoulder-elbow active exoskeleton. The controller predicted the direction of reaching movements across eight possible targets positioned on a 25 cm circumference, achieving an average accuracy of 74.12%. Given the geometric structure of the task, we introduced a macro-direction measure of goodness (MDG) metric that considered both correct predictions and those corresponding to targets adjacent to the true one, resulting in an average performance of 96.63%. Moreover, RC-ID outperformed a kinematics-only benchmark before kinematic onset and surpassed an EMG-only benchmark during the later phases of the reaching movement execution. Finally, effects of assistance were assessed by evaluating the variation of muscular activation during exoskeleton-assisted movements, which led to reductions up to -47.4% with respect the activations during unassisted movements.

Keywords: Whole-Body Motion Planning and Control, Optimization and Optimal Control, Motion Control
Abstract: Legged robots possess a unique ability to traverse rough terrains and navigate cluttered environments, making them well-suited for complex, real-world unstructured scenarios. However, such robots have not yet achieved the same level as seen in natural systems. Recently, sampling-based predictive controllers have demonstrated particularly promising results. This paper investigates a sampling-based model predictive strategy combining model predictive path integral (MPPI) with cross-entropy (CE) and covariance matrix adaptation (CMA) methods to generate real-time whole-body motions for legged robots across multiple scenarios. The results show that combining the benefits of MPPI, CE and CMA, namely using model predictive optimized path integral (MPOPI), demonstrates greater sample efficiency, enabling robots to attain superior locomotion results using fewer samples when compared to typical MPPI algorithms. Extensive simulation experiments in multiple scenarios on a quadruped robot show that MPOPI can be used as an anytime control strategy, increasing locomotion capabilities at each iteration.

Keywords: Task and Motion Planning, Multi-Contact Whole-Body Motion Planning and Control, Humanoid Robot Systems
Abstract: This work presents an optimization-based task and motion planning (TAMP) framework that unifies planning for locomotion and manipulation through a shared representation of contact modes. We define symbolic actions as contact mode changes, grounding high-level planning in low-level motion. This enables a unified search that spans task, contact, and motion planning while incorporating whole-body dynamics, as well as all constraints between the robot, the manipulated object, and the environment. Results on a humanoid platform show that our method can generate a broad range of physically consistent loco-manipulation behaviors over long action sequences requiring complex reasoning. To the best of our knowledge, this is the first work that enables the resolution of an integrated TAMP formulation with fully acyclic planning and whole body dynamics with actuation constraints for the humanoid loco-manipulation problem.

Keywords: Legged Robots, Optimization and Optimal Control, Multi-Contact Whole-Body Motion Planning and Control
Abstract: Automatically generating agile whole-body motions for legged and humanoid robots remains a fundamental challenge in robotics. While numerous trajectory optimization approaches have been proposed, there is no clear guideline on how the choice of floating-base space parameterization affects performance, especially for agile behaviors involving complex contact dynamics. In this paper, we present a comparative study of different parameterizations for direct transcription-based trajectory optimization of agile motions in legged systems. We systematically evaluate several common choices under identical optimization settings to ensure a fair comparison. Furthermore, we introduce a novel formulation based on the tangent space of SE(3) for representing the robot's floating-base pose, which, to our knowledge, has not received attention from the literature. This approach enables the use of mature off-the-shelf numerical solvers without requiring specialized manifold optimization techniques. We hope that our experiments and analysis will provide meaningful insights for selecting the appropriate floating-based representation for agile whole-body motion generation.

Keywords: Imitation Learning, Bimanual Manipulation, Perception for Grasping and Manipulation
Abstract: We investigate the integration of attention maps from a pre-trained Vision Transformer into voxel representations to enhance bimanual robotic manipulation. Specifically, we extract attention maps from DINOv2, a self-supervised ViT model, and interpret them as pixel-level saliency scores over RGB images. These maps are lifted into a 3D voxel grid, resulting in voxel-level semantic cues that are incorporated into a behavior cloning policy. When integrated into a state-of-the-art voxel-based policy, our attention-guided featurization yields an average absolute improvement of 8.2% and a relative gain of 21.9% across all tasks in the RLBench bimanual benchmark.

Keywords: Modeling and Simulating Humans, Human and Humanoid Motion Analysis and Synthesis, Software Tools for Robot Programming
Abstract: Humanoid control remains a central challenge in robotics, as directly replicating full human degrees of freedom(DoFs) incurs significant computational costs and is constrained by mechanical and energetic limitations. To address this, we investigate the hierarchical structure of human movement, where a small set of dominant synergies govern global joint-space coordination. We introduce SynSculptor, a humanoid motion editing framework that leverages synergy-based task-space mapping for training-free, physically plausible human motion scripting. We collect 3+ hours of motion capture data across 20 individuals. We extract primary postural synergies via eigenvalue decomposition of momentum-parsed joint velocity trajectories, constructing a style-conditioned synergy library for free-space motion generation. Biomechanical simulation in OpenSim reveal that the human musculoskeletal model achieves 3.3× greater kinematic efficiency compared to direct motion mapping onto a Supraped HRP4c humanoid. Our extracted kinematic synergy primitives reproduce full-body poses with 96% fidelity using only three joint-space components. By projecting motions into the learned synergy subspace, our reconstructions reproduce kinematic and energetic metrics—∆P, ∆KE, and foot-sliding ratio. We embed synergy-conditioned representations in a motion-language transformer via null-space projections to facilitate personalized humanoid control without additional training for energy-efficient, style-controllable humanoid motion.

Keywords: Bimanual Manipulation, Telerobotics and Teleoperation, Humanoid Robot Systems
Abstract: State of the art robotic systems still struggle to accomplish complex tasks, such as bimanual manipulation, in unpredictable environments due to limitations in sensing, adaptability, and control - leading to artificially restricted autonomous system performance. This research addresses this limitation by introducing a novel metric for quantifying antagonistic internal forces during teleoperated bimanual manipulation, conveyed through minimalistic vibrotactile feedback (VF) to the operator. Experimental tests demonstrated that integrating the proposed metric through VF improved internal force regulation, reducing average internal forces by approximately 39.5% and decreasing force variability by 52.1% compared to trials without feedback. Introducing VF exhibited an increased number of occurrences of payload drops – indicative of operators pushing closer to operational limits. These findings confirm the proposed metric delivered through minimalistic VF substantially enhances teleoperated manipulation, providing a promising avenue for safer, more intuitive robotic operations in dynamic and uncertain environments

Keywords: Force and Tactile Sensing, Reactive and Sensor-Based Planning, Perception for Grasping and Manipulation
Abstract: A typical trade-off exists in the design of current robotic tactile sensing systems: achieving a receptor density comparable to that of human skin often comes at the cost of increased fabrication complexity and reduced device robustness. Super-resolution algorithms offer a promising solution by enabling lager bulk but robust sensing units (e.g., MEMS barometers) to attain enhanced spatial resolution, thereby effectively reducing the hardware complexity of tactile sensing systems. However, dynamic tactile signal super-resolution remains challenging due to the complexity of high-frequency feature extraction. Inspired by the fast-adapting mechanoreceptors in human skin, we propose a biomimetic fingertip featuring a topologically optimized array of 7-MEMS microphones. This design enables precise localization within overlapping receptive fields by leveraging improved signal processing method and an Adaptive Spatio-Temporal Graph Convolutional Network (AST-GCN), resulting in a tactile localization accuracy of 0.5 mm, corresponding to a super-resolution factor of 43. Experimental validation through rolling-sphere trajectory tracking confirms the system's ability to achieve submillimeter precision. The dynamic tactile sensing system provides a robust solution for high-accuracy dynamic manipulation in robotics, particularly in vision-obstructed scenarios requiring fine tactile feedback.

Keywords: Mechanism Design, Education Robotics, Motion Control
Abstract: This paper presents Forte, a fully 3D-printable, 6- DoF robotic arm designed to achieve near industrial-grade performance – 0.63 kg payload, 0.467 m reach, and sub-millimeter repeatability – at a material cost under 215. As an accessible robot for broad applications across classroom education to AI experiments, Forte pushes forward the performance limitations of existing low-cost educational arms. We introduce a cost-effective mechanical design that combines capstan-based cable drives, timing belts, simple tensioning mechanisms, and lightweight 3D-printed structures, along with topology optimization for structural stiffness. Through careful drivetrain engineering, we minimize backlash and maintain control fidelity without relying on high-power electronics or expensive manufacturing processes. Experimental validation demonstrates that Forte achieves high repeatability and load capacity, offering a compelling robotic platform for both classroom instruction and advanced robotics research.

Keywords: Telerobotics and Teleoperation, Reinforcement Learning, Dual Arm Manipulation
Abstract: Full-sized humanoid robot capabilities have grown exponentially in recent years, aiming towards general-purpose deployment in human environments. A popular control method used by manufacturers utilizes Virtual Reality for upper-body teleoperation and Reinforcement Learning for lower-body balance and locomotion control. As a result, a single remote operator can see, manipulate, and navigate about a real, distant physical environment. This powerful control stack is often relegated to expensive full-sized robots, many of which are inaccessible to the research community. Miniature humanoids are more prevalent, but employ less biomimicry in their design (e.g. fewer sensors, Degrees of Freedom, etc) and lack similar developments. This paper describes a compliant full-body telepresence control stack developed from the ground up for miniature humanoids. Framework experimentation on ROBOTIS OP3 hardware showcases walking at speeds up to 0.45 m/s independent of arm motions. Tele-loco-manipulation is demonstrated via a cube relocation experiment with an expert human operator. On average, the teleoperated system moved 2 different 40 g cubes within 10 mins, walking a total distance of 5 m. Overall, the developed system shows potential for miniature humanoid tele-loco-manipulation.

Keywords: Force Control, Compliance and Impedance Control, Motion Control
Abstract: This paper presents the Inertia-Integrated Wrench Observer (IIWO) for highly sensitive zero-wrench control. IIWO embeds task-space inertia into a disturbance observer to decouple internal dynamic reactions from true external wrenches, preserving sensitivity to small forces and moments. Experimental peg-in-hole insertion results show that IIWO achieves deeper, more compliant insertion with minimal residual wrench, outperforming both a conventional FOB and a PD controller baseline.

Keywords: Humanoid and Bipedal Locomotion, Compliance and Impedance Control, Mechanism Design
Abstract: This paper proposes a humanoid robot foot designed with aluminum and soft materials, which can be easily implemented. The proposed foot effectively reduces ground reaction force (GRF) and enhances walking stability through a mechanical filtering effect and improved performance of an admittance controller. Experimental validation using the DRC-Hubo+ platform demonstrates that the proposed foot reduces torque and GRF tracking errors compared to the conventional foot and improves Zero-Moment-Point (ZMP) tracking performance.

Keywords: Bimanual Manipulation, Dual Arm Manipulation, Task and Motion Planning
Abstract: Efficient, human-like object transfer between cooperating robots demands both spatial precision and tight temporal coordination. Existing approaches treat these requirements in isolation or rely on pre-computed trajectories that fail when obstacles appear, degrading performance and introducing desynchronization. This paper introduces a unified dynamical systems framework that transitions each arm from independent asynchronous motion to coupled synchronous coordination. An obstacle-aware modulation layer steers end-effectors smoothly around obstacles, thus avoiding them, without re-planning. Experiments on a humanoid platform and on traditional manipulators show seamless handovers that remain stable despite obstructions, always preserving spatial and temporal synchrony.

Keywords: Force and Tactile Sensing, Soft Sensors and Actuators, Soft Robot Materials and Design
Abstract: We introduce Flexitact, a conformable tactile sensor array concept fabricated via multi-material 3D printing on a TPU95A substrate using electrically conductive PLA. Each sensing unit ("flexel") is proposed as an interdigitated, deformable capacitor whose capacitance varies with local pressure. We formulate an RC time-to-threshold readout model, derive expected scaling of scan time with array size, and present a peak-based multi-contact differentiation framework for separating one strong press from multiple light touches. This work focuses on the theoretical design, sensing model, and algorithmic pipeline for Flexitact; experimental construction and quantitative validation are outlined as planned evaluation.

Keywords: Engineering for Robotic Systems, Methods and Tools for Robot System Design
Abstract: Robotic hands designed for general-purpose tasks have demonstrated remarkable capabilities; however, they continue to encounter challenges in real-world industrial scenarios. The requirements for industrial automation are distributed across a wide range. In the pursuit of meeting all criteria, often uneconomical concepts are designed, or crucial requirements are dropped. Instead of following the general-purpose approach we propose a method to set up use-case dependent, multi-purpose end effectors. We identify and cluster industrial applications and use a platform concept to configure suitable anthropomorphic multi-purpose end effectors. One preliminary design derived from the platform for assembly processes is presented. This design allows to be easily configured to fit to the defined tasks and objects within the assembly application. This approach contributes to the transformation from robotic hands in labs to economical usage in industry.

Keywords: Wearable Robotics, Rehabilitation Robotics, Soft Sensors and Actuators
Abstract: Human movement kinematics analysis is essential in various applications such as medical, sports and robotics. An inter-segment pose estimation sensor was proposed in our previous. study. In this study, human body kinematics, especially around knee, was reconstructed using the sensor.

Keywords: Wearable Robotics, Physically Assistive Devices, Mechanism Design
Abstract: This study proposes a powered ankle-foot orthosis for individuals with post-stroke hemiparesis that avoids forced movement constraints and eliminates the need for external anchoring. The device comprises an insole, a shank fixation part, and a flexible hinge, allowing the insole’s ground contact to serve as an anchor while preserving natural joint motion. The system, including sensors, wearable components, and controller, has been implemented and is undergoing clinical evaluation.

This mechanism is expected to enable more effective gait rehabilitation using assistive robotics.

Keywords: Compliant Joints and Mechanisms, Mechanism Design, Soft Sensors and Actuators
Abstract: This paper presents a novel spring design that enhances the Capacity-to-Stiffness without increasing actuator volume or introducing morphological complexity. The proposed design employs multiple cantilever beams arranged in parallel, enabling the stiffness to be preserved while increasing capacity through beam thinning. Analytical expressions for stiffness and capacity are derived based on cantilever beam theory, and the theoretical relationship between the number of beams and the Capacity-to-Stiffness is established. To validate the proposed methodology, Finite Element Analysis (FEA) was performed on ten spring samples with increasing beam numbers. The results confirm that higher beam counts lead to improved capacity under fixed stiffness conditions.