Papers

Ensuring safety of learning-enabled robotic manipulation across diverse embodiments and tasks still requires significant manual engineering. For collision avoidance, existing approaches typically rely on embodiment-specific learned safety approximations or simplified robot models together with stopping-based fallback controllers that compromise performance. In this paper, we propose an embodiment-agnostic safeguarding approach, X-Safe, which provides formal probabilistic guarantees for collision avoidance without model simplifications or learned safety functions. By operating in the configuration space, our safeguarding method works with different types of robot policies. X-Safe reasons locally around the current configuration to establish safe regions and uses action masking to restrict the policy's support to these safe regions. We demonstrate X-Safe on two embodiments and three policies, both in simulation and on hardware. We observe less than half of the performance degradation compared to state-of-the-art safeguarding, and empirically corroborate our formal collision avoidance guarantees.

Cooperative aerial transport requires controllers that respect nonlinear manifold geometry, operate without centralized coordination, and respect operational safety constraints. To address these demands, we present GPAC, a four-layer hierarchical architecture that enables $N$ quadrotors to transport a cable-suspended payload without a central coordinator or by exchanging cable states or adaptive parameters. The key insight is implicit coordination: each quadrotor independently estimates its effective load share from local cable measurements, so combined forces converge to the correct total, even without knowledge of $N$ or the payload mass; the payload position is reconstructed locally from each agent's own cable geometry, and the only inter-agent communication is a low-rate neighbor-position broadcast for collision avoidance. GPAC operates directly on the full nonlinear configuration manifold and integrates geometric position and attitude control, anti-swing regulation, an extended-state observer for wind rejection, concurrent learning-based mass estimation without persistent excitation, and a priority-ordered control barrier function (CBF)-inspired safety filter that reduces operational risk, with input-to-state safety (ISSf) margins that hold exactly under single-constraint activation. A compatibility result shows that the filter's force modifications keep the desired attitude within the almost-global stability region of the $\mathrm{SO}(3)$ attitude controller. Finally, high-fidelity simulation with flexible cables, onboard sensor fusion, and wind turbulence – with all control and estimation loops closed through the estimator – yields a mean payload-tracking RMSE of 33.8 cm (2.8% coefficient of variation over 13 seeds) at a low per-agent computational cost.

Tactile sensing enables robots to perceive rich contact information at the grasp, supporting tasks such as object recognition, in-hand pose estimation, and slip detection. However, in many tool-mediated manipulation tasks, the interaction that determines task success occurs at the tool tip, away from the tactile sensor, making direct sensing of tool-environment contact difficult, particularly when the contact moves during interaction. In this work, we reconstruct the trajectory of extrinsic tool-tip contact using tactile sensing and robot proprioception. We formulate tool-tip trajectory reconstruction as a geometric inference problem under a single-point contact assumption. Our method first estimates the global tool-tip contact location from a calibration segment designed to approximate fixed-point behavior, and then reconstructs the full trajectory by composing relative tool motion estimated from tactile marker observations under continuous contact. Across n=51 trials with multiple trajectories, tools, wrist poses, and grasp configurations, the proposed pipeline achieves a trajectory RMSE of 8.59 +/- 2.41 mm in the world frame and a shape RMSE of 5.96 +/- 1.16 mm, while operating online at 14.00 +/- 4.11 Hz. Overall, the results show that extrinsic tool-tip trajectory geometry can be recovered consistently from grasp-level tactile sensing, with trajectory shape remaining stable across variations in tools, wrist poses, and grasp configurations.

Whole-body humanoid loco-manipulation requires coordinating the robot's entire kinematic chain. However, most existing systems typically decouple the upper and lower bodies into separate controllers, limiting such coordination and yielding behaviors similar to those of a wheeled dual-arm platform. In this paper, we ask what it takes to build a whole-body native vision-language-action (VLA) model that maps language and pixels directly to all of the humanoid's degrees of freedom. We conduct a systematic empirical study organized as a roadmap of one-variable-at-a-time experiments across three phases: whole-body teleoperation, VLA model design, and heterogeneous co-training. Our study yields several intriguing findings: a joint-based whole-body teleoperation interface outperforms alternatives that only partially expose the humanoid's degrees of freedom; a VLA pretrained on static and wheeled dual-arm platforms transfers surprisingly well to a humanoid's full action space; and co-training with HuMI, the humanoid analog of UMI, extends the policy to new objects and instructions without additional whole-body teleoperation on those targets. Following this roadmap yields OpenHLM, an open-source recipe for whole-body humanoid loco-manipulation. In a challenging long-horizon task that spans a wide vertical range of the humanoid, OpenHLM outperforms two state-of-the-art humanoid VLA baselines (GR00T N1.6 and $\Psi_0$) using less than half the total demonstration time. Our code, training data, and model checkpoints are available at [https://openhlm-project.github.io/].

This paper presents a full nonlinear constrained dynamic model of MonoRollBot, a novel 3-DoF spherical rolling robot driven by a single motor, a lead-screw transmission, and a spring-coupled internal moving mass, together with motion analysis of its behavior. To the best of our knowledge, this is one of the first full nonlinear nonholonomic models reported for a mono-actuated, super-underactuated spherical rolling robot of this kind. Because rolling without slipping is nonholonomic, the dynamics are derived using the Lagrange–d'Alembert formulation, with the lead-screw relation imposed as a holonomic constraint and the rolling condition imposed in Pfaffian form. The formulation retains the complete generalized coordinates of shell translation, shell attitude, screw travel, nut rotation, and radial mass motion. Simulations and representative motion studies show qualitative agreement with prototype behavior and reveal how gravity, compliance, and inertia jointly shape the locomotion and motion capabilities of this strongly underactuated robot. The resulting model also provides a mechanically consistent basis for future state estimation and hybrid controller design for this nonholonomic mono-actuated rolling robot.

Reinforcement learning (RL) is a powerful and convenient tool to modernize controller design. In this work, we study the zero-shot transfer of RL-based control policies from simulation to hardware for cart-pole swing-up and stabilization. The two policies are trained independently, and the handoff is implemented in Simulink via switching logic. We apply a first-order action smoothing filter to prevent hardware damage from high-frequency oscillatory actuation. Pairing this bandwidth-aware filtering with sensitivity-guided domain randomization (DR) and a simple linear curriculum learning (CL) schedule, we obtain a swing-up policy that in all of our experiments injects sufficient energy for handoff into the stabilizer's region of attraction. The stabilization policy rejects disturbances within the tested range, and the swing-up policy can re-engage after larger perturbations and restores the pendulum to the inverted position.

Planning contact-rich whole-arm manipulation is challenging because interactions that involve extended robot geometry give rise to complex contact dynamics that are difficult to model accurately. This creates a need for planning principles that do not rely heavily on precise contact models. Caging offers one such geometric notion of robustness to modeling inaccuracy by restricting object escape through geometrically enclosing the object. However, existing caging formulations are difficult to incorporate into continuous optimization-based manipulation planning. We reformulate caging as a minimum-time escape problem in which the object seeks to leave an enclosing robot geometry in the shortest time. This yields a continuous escape-time field that measures the robot's enclosure quality and we show it satisfies an eikonal equation. We therefore can approximate this field using a physics-informed neural network, producing a smooth differentiable representation that can be embedded directly into manipulation planning. The resulting objective supports whole-arm manipulation planning to favor robot configurations resisting object escape. This improves the manipulation robustness to contact model mismatch, thus enabling planning with simplified contact models, including quasi-dynamic approximations and simplified object geometry. Across simulation and real-world experiments, we show improved robustness to disturbances and contact-model mismatch relative to baselines. These results suggest that geometric enclosure can serve as a practical robustness primitive for whole-arm manipulation. A supplementary video, which includes an intuitive overview of our method and experiment video results, is available on our project webpage.

We present RoboLineage, an agent-native data lifecycle governance system for robot policy iteration. Modern robot policies improve through repeated data collection, review, retraining, evaluation, and release decisions, but the evidence connecting these steps is often scattered across local tools, scripts, and expert memory. RoboLineage makes this lifecycle explicit by representing rollouts, reviews, dataset decisions, training runs, policy metadata, evaluations, deployment recommendations, and next-collection plans as typed lineage artifacts. Agents interpret embodied rollout evidence, adapt accepted data to existing training stacks, maintain data health, and summarize cross-iteration state under explicit artifact boundaries. In real-robot manipulation workflows, RoboLineage makes routine policy iteration faster and more auditable while maintaining downstream policy performance. We open source RoboLineage as a lightweight lifecycle layer for different robot embodiments and training families. Project page: https://robolineage.github.io/

Scaling dexterous manipulation requires generalization across objects, scenes, and tasks, yet existing data sources face a trade-off between scale and scene/embodiment alignment: teleoperation data is well aligned with robot deployment but expensive to collect; simulation is scalable but limited by the sim-to-real gap; and real egocentric videos scale effectively but remain misaligned with robot deployment. We propose Wh0, a framework that uses generative video world models as scalable and controllable sources of egocentric human-hand manipulation data to unlock the manipulation capabilities of pretrained dexterous VLA models. Conditioned on language, objects, and scenes, Wh0 uses a generative world model to produce WM-H, a 50k-episode dataset of egocentric human-object interaction videos. Wh0 then converts the generated videos into robot-trainable supervision through hand motion reconstruction and visual editing. Co-trained with a limited amount of real robot data, WM-H adapts pretrained VLA models to dexterous manipulation deployment. Across 18 real-world dexterous manipulation tasks, compared with a model post-trained only on robot data, Wh0 improves zero-shot success on unseen tasks from 8.3% to 38.9%. Ablation studies further show that scalable generation and scene/embodiment alignment are key drivers of performance gains. Videos and open-source code can be found on our project website: https://chenyt31.github.io/wh0.github.io/.

The Jansen linkage is a single-degree-of-freedom planar leg mechanism whose eleven "holy numbers" were evolved by Theo Jansen to optimize the foot-path gait alone, with no regard for the wear of its revolute joints. This paper introduces a durability objective into the design of the Jansen leg. A parametric forward-kinematic model (two-circle-intersection solver), an inverse-dynamic model (constraint-Jacobian / Lagrange-multiplier formulation of a seven-body, ten-joint system, independently cross-verified by a reduced-DOF energy method), and an Archard wear model are coupled to evaluate, for any set of link lengths, both gait quality and the per-cycle sliding wear at every pin. Because the wear is computed on ideal, clearance-free revolute joints, the resulting wear figures are a relative comparative ranking rather than an absolute life prediction. A bi-objective problem – composite gait error versus total joint wear, subject to step-length, ground-clearance, duty-factor and assembly constraints – is solved with NSGA-II. Under the adopted gait metric the classical Jansen design is Pareto-dominated: for a representative design, link-length adjustments within +/-29% simultaneously flatten the stance (-28%), smooth the stance velocity (-58%) and reduce total joint wear by ~56%. A sensitivity study shows the wear advantage is robust across a crank-speed x payload envelope (48%-56%) and identifies the link lengths that most strongly govern wear. A variance-based global (Sobol) analysis confirms that two link lengths dominate the wear variance, and a Monte-Carlo manufacturing-tolerance study shows the wear advantage degrades gracefully under realistic fabrication error. The framework provides a practical route to longer-lived walking linkages and a baseline for future wear-clearance-impact coupled studies.

Deformable linear objects (DLOs) such as wires, cables, and ropes are common in robotic manipulation tasks, yet simulating them with both visual realism and physical accuracy remains challenging. Existing visual simulation methods typically rely on procedural geometric primitives that lack physically grounded deformation behavior, while physics-based approaches with robot learning support often approximate DLOs as rigid-link chains or generic soft bodies, failing to accurately capture the bending, twisting, and shear mechanics of slender elastic structures. In this work, we introduce DeformX, a co-simulation framework that integrates a dedicated Cosserat rod physics engine with NVIDIA Isaac Sim, enabling DLO simulations that are both physically faithful and visually realistic. Our Cosserat rod engine simulates the dynamics and self-collisions of DLOs, and contact interactions with arbitrary free-form meshes. To achieve high-fidelity visualization, we employ mesh skinning to map discrete rod deformations onto imported CAD models. To the best of our knowledge, DeformX is the one of the first frameworks for DLO simulation that unifies realistic visualization, principled physics, and compatibility with robot learning pipelines. We demonstrate its versatility across synthetic data generation and policy learning for DLO manipulation, and validate visual and physical fidelity through comparisons against real-world experiments. Notably, fine-tuning Segment Anything Model 3 (SAM3) on DeformX-generated data yields a 10.2% mAP@75 improvement in real-image wire segmentation, and a rope-swinging policy trained entirely in DeformX achieves a mean target-hitting error of 6.6 cm on a UR5e manipulator in real-world trials, highlighting its strong sim-to-real transfer capability.

Generalizing manipulation policies across robot embodiments remains difficult because standard policies tie the interactions needed for a task to the actions of a particular robot. We study zero-shot cross-embodiment manipulation, where a policy trained on source embodiments must be deployed on a structurally different target embodiment without additional task demonstrations. We introduce Kinematic Interaction Transfer across Embodiments (KITE), which decouples manipulation into embodiment-agnostic interaction prediction and embodiment-specific kinematic decoding, connected through a learned representation of interaction intent based on contact patterns. An interaction policy is trained to predict the intent using demonstration data only collected on a source robot, and shared across morphologically different target robots. A kinematic decoder realizes these intents by learning from synthetic data generated from each embodiment's kinematic model. Through such decoupling, KITE can robustly transfer learned skills to a new embodiment given only its kinematic model, without recollecting any task-specific demonstration data. We evaluate KITE on four manipulation tasks in simulation and the real world spanning transfer between parallel grippers, dexterous hands, and composite embodiments. KITE successfully achieves zero-shot transfer to structurally different target embodiments, and consistently outperforms state-of-the-art baselines. Project website: https://kite-manip.github.io/.

Per-scene 3D Gaussian Splatting (3DGS) enables high-fidelity rendering, but practical robotic and AR scene capture pipelines often depend on external geometric initialization (e.g., SfM point clouds or depth estimates), which can be slow and brittle in on-site deployment. We present ACEsplat, a fast per-scene optimization framework that reconstructs 3D Gaussian representations from RGB images and camera poses only, without requiring external 3D priors (e.g., precomputed SfM models or supervised depth maps). ACEsplat uses a two-stage pipeline: (1) a self-supervised scene coordinate regression (SCR) module builds an internal geometry prior within 4–5 minutes; (2) SCR features and coordinate priors are fused by a lightweight Gaussian initialization head, followed by per-scene 3DGS optimization. On static-view rendering, ACEsplat achieves 29.11 dB PSNR on Wayspots with real-time SLAM poses and 33.20 dB on Cambridge Landmarks with SfM-refined poses. On RealEstate10K sparse-view novel view synthesis, it achieves competitive image fidelity under a challenging 2-view setting. ACEsplat completes scene-specific SCR mapping and 3DGS reconstruction within 15–25 minutes on a single GPU, making it a practical RGB+pose-only solution for rapid scene setup in robotics and mixed-reality applications.

This paper proposes an energy-recycling rimless wheel with spring-clutch legs. The proposed mechanism uses a lockable clutch to store part of the impact-induced elastic energy after foot contact and reinject it in the next gait cycle. First, we develop a hybrid dynamic model of the energy-recycling rimless wheel. Second, numerical simulations are used to examine the dynamics, local stability of periodic gaits, and the Cost of Transport (CoT) of the proposed mechanism. The simulation results show that the proposed mechanism reduces the CoT by up to 16.13% compared with a benchmark viscoelastic-legged rimless wheel with telescopic spring-damper legs. Compared with the rigid rimless wheel, the viscoelastic-legged and energy-recycling models reduce the CoT by more than 50%. The energy-recycling model also maintains locally stable periodic gaits over the tested slope and stiffness ranges. Finally, prototype experiments on an inclined plane are conducted to examine the feasibility of the proposed mechanism. The experimental results show that the proposed rimless wheel achieves passive walking on a shallow 1{\deg} slope, corresponding to a CoT of approximately 0.02. These results suggest that the proposed spring-clutch mechanism can improve the simulated walking efficiency of the energy-recycling rimless wheel, while the prototype experiments support the feasibility of passive walking with the mechanism.

Simulation-to-reality transfer, often called sim-to-real transfer, is a central challenge in robot learning. Yet, the tradeoff between measuring a system more accurately and training over a broader range of simulated dynamics is still poorly understood. In this work, we focused on the allocation of real-robot measurement time between system identification and domain randomization. We studied this tradeoff in a controlled sim-to-sim pendulum setting, where a hidden-parameter model stands in for the physical robot, and the experiment sweeps identification rollouts against the width of the randomization distribution. Across the reality gaps and noise levels we tested, the measurement budget did most of the work. A small number of identification rollouts closed most of the transfer gap, and once any real data was available, policies performed best when trained at the estimated parameters rather than over a widened randomization band. Broad randomization that contained the true system still did not substitute for measurement. These results hold in a benign regime where the dynamics are identifiable and only two parameters are unknown, so structural model mismatch remains the setting where randomization breadth may become more valuable. Overall, our results suggest that sim-to-real pipelines should first measure the parameters they can and reserve randomization for the uncertainty that remains.

Maneuver coordination is a key enabler of connected and automated driving, allowing vehicles to negotiate and execute maneuvers that would otherwise be difficult, inefficient or unsafe. Existing approaches and use cases typically assume coordination with a single predefined target vehicle, which limits the number of coordination opportunities. This paper introduces a maneuver coordination approach based on multi-target selection, which allows a vehicle to identify and select among multiple potential coordination vehicles for a given maneuver. Multi-target maneuver coordination does not require modifications to the maneuver execution logic or to the underlying coordination protocol. Instead, it extends the decision-making process preceding coordination, enabling vehicles to exploit a broader set of feasible cooperative interactions. Results show that multi-target maneuver coordination significantly increases triggered and successfully executed coordinations while maintaining a low computational cost, as the proposed approach achieves these gains without requiring the analysis of a large number of potential target vehicles. These improvements preserve coordination success rates while enabling earlier maneuver initiation.

Monocular and RGB-D visual-inertial SLAM systems remain susceptible to limited field of view, sensor-specific failure modes, and unreliable cross-session relocalization. To address these issues, we present GeoFlow-SLAM++, a tightly coupled multi-camera visual-inertial SLAM system that extends GeoFlow-SLAM from a single RGB-D sensor to a calibrated multi-camera rig with a unified body-centric formulation. Within this multi-camera framework, GeoFlow-SLAM++ supports two interchangeable visual front-ends: a conventional ORB front-end and a neural network feature (NN-Feature) front-end built on SuperPoint and LightGlue. The system unifies tracking, mapping, and relocalization on a shared body state, and combines multi-camera reprojection constraints, IMU pre-integration, cross-view place recognition, and dual-stream optical flow/NN-Feature tracking for robust localization. As an optional extension, the system can further incorporate cross-view-consistent pseudo-depth predictions from RGB images as auxiliary geometric constraints. We evaluate GeoFlow-SLAM++ on EuRoC, OpenLORIS, TUM, Hilti, and a self-collected handheld multi-camera dataset. Results show that the NN-Feature front-end improves robustness in appearance-challenging scenarios, the multi-camera formulation achieves competitive localization accuracy on Hilti, and the unified cross-view relocalization design reaches LiDAR-comparable performance on the handheld dataset.

Tailless flapping-wing micro-aerial vehicles (FWMAVs) mimic the impressive flight performance of hummingbirds, utilising unsteady aerodynamic effects. However, existing designs are still limited and purpose-built with a restricted flight envelope and poor endurance. We therefore propose an adaptable tiltwing framework enabling bioinspired aerial robots to switch between hovering flight, high-speed directional flight, and energy-efficient gliding flight. The proposed framework utilises thrust vectoring with a wide actuation range via two fully independent propulsion units, each flapping a single wing, for effective control and enhanced manoeuvrability. For this, we developed a hybrid Scotch-yoke-based flapping mechanism that ensures a symmetric motion profile with a modular design guaranteeing an arbitrarily wide flapping angle to exploit the lift-enhancing clap-and-fling effect. Additionally, we implemented a passive wing-rotation mechanism, which, in combination with our dual-wing thrust-vectoring approach, allows unprecedented wing-design freedom, unlocking potential for precise optimisation. A contactless leading-edge tracking sensor provides accurate feedback on the wing's orientation and, in the gliding mode, enables dihedral-angle control, augmenting the active wing-pitch control. Extensive testing of a propulsion unit was conducted with a six-axis force/torque sensor, demonstrating the flapping mechanism's performance while optimising transmission efficiency and the passive wing-pitch mechanism. At full throttle, the average lift force generated by a single wing, flapping with a 188{\deg} amplitude, was 21.1 gf for a small 3.1 g 1S BLDC motor. Additional tests covering the full range of the wide-angle tilting capability showed an effective thrust-vectoring control architecture with a linear and symmetric response curve of the moments generated.

Sit-to-stand (STS) transitions impose significant joint-loading demands on elderly individuals, making them a primary target for lower-limb exoskeleton assistance. However, accurate trajectory tracking during STS is challenging due to complex, time-varying human exoskeleton interaction dynamics and inter-subject variability that render model-based control approaches difficult to apply in practice. This paper presents an intelligent model free adaptive backstepping control strategy for a bilateral lower-limb exoskeleton during STS motion. The proposed controller design uses an ultra-local second-order model to avoid explicit system identification, while a Gaussian radial basis function (RBF) neural network estimates the unknown lumped dynamics online. To further improve phase-aware tracking performance, a Twin Delayed Deep Deterministic Policy Gradient (TD3) reinforcement learning agent is integrated as a supervisory gain scheduler that adaptively adjusts controller gains across the distinct phases of STS motion. The proposed controller is evaluated through co-simulation in MATLAB/Simulink and Simscape Multibody using OpenSim-derived reference trajectories and benchmarked against state-of-the-art controllers. Results demonstrate that the proposed controller achieves the lowest average RMSE of 0.078 degree across all joints, representing improvements of 60.2%, 54.4%, 48.7%, and 42.6% over proportional integral derivative (PID), model-free adaptive control (MFAC), linear quadratic regulator (LQR), and sliding-mode control (SMC), respectively. TD3 integration further reduces tracking error by 35%, 33%, and 79% at the hip, knee, and ankle joints compared to the standalone RBF-MFAC baseline. These results demonstrate the effectiveness and robustness of the proposed controller design for assistive exoskeleton control during STS transitions.

Reinforcement learning for robot manipulation is often bottlenecked by reward design, especially in long-horizon tasks: sparse success rewards provide weak supervision, while hand-crafted dense rewards are tedious to design and generalize poorly across tasks. Progress-based reward models offer a promising alternative by estimating how far an observation has advanced toward task completion, but existing approaches often require task-specific demonstrations or progress labels, and can assign high rewards to visually plausible but physically incorrect states. We introduce the Reference-Anchored Reward Model (RARM), a lightweight visual comparator that converts a single successful demonstration into a dense, progress-aware reward. RARM is trained once on general-purpose videos with a contrastive temporal objective, requiring no robot-specific data, task-specific reward labels, or per-task reward engineering. At deployment, RARM matches rollout clips to reference clips and rewards only confident forward progress, suppressing uncertain matches that may otherwise produce false-positive rewards. Across 9 simulated manipulation tasks from LIBERO and MetaWorld and 4 real-world tasks, RARM achieves the best overall success rates in subsequent RL training, with particularly large gains on long-horizon tasks such as cloth folding, where unreliable progress estimates are especially harmful.

Diffusion models excel at capturing multi-modal action distributions in robot imitation learning. However, in multi-task and long-horizon scenarios, monolithic architectures lack structural generalization capabilities, suffering from gradient conflicts between distinct semantic sub-stages. While pure data-driven Mixture-of-Experts (MoE) methods introduce labor division, they frequently trigger routing collapse, and instantiating full-scale experts causes parameter explosion and high expansion costs. To address these issues, we propose Concept-prior Routed Diffusion Experts (CoRDE), a structure-guided variational distillation framework. CoRDE extracts semantic distributions from a frozen concept encoder to guide the variational posterior responsibility via a learnable soft mapping matrix. This mechanism introduces an entropy-controlled responsibility inference process that encourages confident routing under reliable semantic predictions while preserving the stochastic diffusion term for behavioral diversity. To overcome parameter inflation, CoRDE employs a parameter-efficient expert pool using Low-Rank Adaptation (LoRA) on a shared frozen backbone. Theoretical analysis shows that the mixture score discrepancy is bounded by responsibility-weighted local expert errors, supporting high-fidelity generation under low-rank expert adaptation. Empirical evaluations confirm that, compared to existing baselines, CoRDE systematically reduces routing collapse, forming robust, semantically aligned expert allocations while achieving superior action quality and incremental learning efficiency.

Goal-directed eye movements are a fundamental component of visuomotor control, enabling humans to anticipate and guide their actions. For this reason, they are increasingly used in human-robot interaction to estimate users' goals. However, during manipulation, fixations may reflect either an intended future action or the need to visually monitor the robotic proxy due to altered embodiment. How predictive and monitoring-related gaze are organized across the different phases of a constrained robot-mediated manipulation remains unclear. Here we address this question by investigating gaze behavior during goal-directed telemanipulation to characterize how visuomotor control adapts to altered embodiment in a multi-stage task. Our findings show that gaze remains strongly aligned with task goals, preserving its predictive role even during robot-mediated manipulation. At the same time, gaze frequently alternates between the robotic end-effector and the manipulated object, revealing increasing online monitoring. The presence and geometry of obstacles modulate the timing and distribution of these fixations, delaying attention to the final target when intermediate constraints become more demanding. These findings show that predictive gaze is not lost under altered embodiment but reorganized in response to changes in sensory feedback and control demands. More broadly, they highlight the flexibility of the human visuomotor system when the natural sensorimotor coupling is disrupted and suggest that gaze should be interpreted contextually rather than treating every fixation as direct evidence of user intention in human-robot interaction.

Co-design of legged robots with elastic elements is challenging due to the non-differentiability of contact dynamics and mechanism engagement. This paper presents SurGE, a framework that computes surrogate gradients of the design objective through a differentiable pipeline consisting of a kinodynamic single-rigid-body (Kino-SRB) model and a design-aware control policy, and injects them into CMA-ES via mean shift with cosine-annealed step decay. On a 4-DOF design space of a hopping robot with unidirectional parallel spring, SurGE achieves 6 times lower cross-seed standard deviation and 18% tighter population concentration compared to vanilla CMA-ES, while matching or improving the best objective. Hardware experiments on a 2D design subspace show that, starting from a hand-tuned initial design, SurGE reduces the design objective by 37.65% on hardware, with the improvement trend identified in simulation transferring consistently to the physical system. SurGE provides the potential to accelerate non-differentiable co-design problems in legged robots via surrogate model gradients.

Manipulation in confined environments, such as threading a manipulator through narrow apertures, remains a fundamental challenge, especially for conventional rigid robots. Hybrid rigid-soft manipulators offer promise but face two compounding planning challenges: backbone shapes feasible in free space become infeasible under environmental contact, and planning rigid and soft segments independently ignores their kinematic coupling. We present THREAD, the first diffusion-based trajectory planner for hybrid manipulation, learning a generative prior over physically realizable backbone trajectories conditioned on local environment geometry, with physics-inspired losses encoding curvature, smoothness, and collision constraints jointly across both segments. Trained in simulation, THREAD achieves 92.4% task success with 5x fewer collisions than the strongest baseline. We show cross-embodiment real-world transfer with minimal online updates, successfully threading through apertures as small as 1.3x the soft segment diameter.

Point-cloud goals provide a direct way to specify dexterous in-hand reorientation: instead of defining an object-specific pose frame or estimating 6D pose at test time, the policy is given the desired 3D geometry of the object. Yet raw point-cloud goal conditioning is poorly conditioned for policy learning. Current and goal clouds are unordered, independently sampled, and often visibility-dependent, so their discrepancy entangles object rotation with permutation, resampling, and unstable correspondence structure. For this reason, prior point-cloud manipulation methods typically add structure outside the representation itself, such as explicit pose or relative-pose inputs, dense flow features, or distillation from privileged teachers. We close this gap by learning a rotation-aware point-cloud embedding whose Euclidean latent distance is calibrated to the SO(3) geodesic error between object orientations. The resulting representation turns current-goal comparison into a smooth control signal, allowing a model-free RL policy to act from current and goal point-cloud embeddings, proprioception, and centroid metadata, without object pose, relative pose, dense flow, or teacher-action supervision. In in-hand reorientation experiments, this interface matches privileged-state and distillation-based baselines while avoiding brittle test-time computation of structured pose or flow inputs. These results suggest that point-cloud goals become practical for this task when the representation, rather than an external module, encodes the task-relevant geometry of rotation. We also show evidence that generic visual point-cloud pretraining is insufficient for such a current-goal comparison because it discards the task-relevant state and preserves only shape features.

Fish achieve efficient swimming across varied speeds through active modulation of their body flexibility. To explore the effects of tunable stiffness on swimming performance, we present a bio-inspired freely swimming fish robot with a rapidly tunable particle-jamming body. This design enables rapid stiffness adjustments with negligible changes in shape or volume, achieving a 54% variation in flexural rigidity across vacuum pressures of 0 to -40 kPa. We visualize the midline of the oscillating body under both low- and high-stiffness conditions, and the comparison confirms that the body curvature varies with stiffness. We further experimentally evaluate the tunable stiffness body's effects on swimming performance using velocity and cost of transport (CoT) measurements obtained via a motion tracking system. Results show that active stiffness tuning is essential for sustaining efficient and high-speed swimming across beating frequencies of 1-3 Hz. At low frequencies (1-1.5 Hz), a softer body (0 kPa) maximizes velocity and minimizes CoT, whereas at high frequencies (2.5-3 Hz), a stiffer body (-40 kPa) delivers superior velocity and reduced transport cost. These findings highlight stiffness modulation as a key strategy for adaptive and efficient propulsion in bio-inspired robotic swimmers.

Lunar micro-rovers under 50 kg are a rapidly growing vehicle class, yet open, benchmarked tools for running optimization across design trades remain scarce. Rover sizing is a highly coupled problem with a feedback loop specific to surface vehicles: a heavier rover sinks deeper into loose regolith, deeper sinkage costs more power, more power requires a larger array, and the array adds mass. We present RoverDevKit, a tradespace exploration toolkit with an open evaluator that couples a closed-form wheel-soil model with bottom-up mass, solar-power, thermal-survival, and traverse models. It is fast enough to serve as the fitness function of a multi-objective genetic algorithm (NSGA-II): one search of 3000 mission evaluations takes under a minute on a single laptop core. We use it to map the trades among distance traveled, vehicle mass, and slope capability on mare, polar, highland, and crater-rim missions. Across those missions the limiting factor changes: energy storage at high latitude, slope traction on loose highland regolith, and traverse range on mare and crater-rim terrain. The wheel-soil model matches measured single-wheel drawbar pull on two independent datasets to within the 20-30% error typically reported for this class of model, and the mass model predicts published 5-50 kg rover masses to 10.5% median absolute error. A check against five published in-class designs places them near the designs the optimizer selects. Re-running the search after applying that wheel-soil comparison error leaves these conclusions unchanged.

Magnetically programmable soft elastomers enable complex shape morphing and locomotion dynamics in small scale soft robots under external magnetic fields. Benefiting from their programmed deformation and wireless actuation capabilities, magnetic soft robots have emerged as promising platforms for targeted drug delivery, especially in human gastrointestinal tract. However, achieving controlled directional liquid cargo release toward desired tissue interface while preserving the encoded shape morphing and locomotion capabilities remain a significant challenge. Here, we report a new design strategy that employs an optimized magnetization profile to enable controlled directional release of aqueous cargo without compromising shape morphing and locomotion capabilities. Magnetic soft robots with a specific spatially distributed magnetization profile allow directional alignment of the release interface with the orientation of the external magnetic field. This orientation control ensures active alignment of the release interface toward the intestinal wall prior to drug release. An interconnected microporous elastomer is embedded within the robot for aqueous cargo storage, while a thin microcrystalline wax layer seals the release opening hole to isolate the stored liquid cargo from external environment during transport. Triggered release is achieved by mechanically rupturing the wax sealing layer under a higher magnitude external magnetic field. Controlled directional flipping, locomotion, and triggered release are decoupled through external magnetic field's direction and strength. The controlled directional release strategy reported here integrates directional targeted liquid cargo release, shape morphing, and locomotion, which establishes the groundwork for target drug delivery in gastrointestinal tract applications.

The capability to encode spatially distinct magnetization patterns within soft materials enables remote control over complex deformations. This characteristic is especially important for microfluidic platforms, where limited dynamic control of channel boundaries and laminar flow conditions usually restrict fluid transport and interactions. The present study introduces a shape-programmable magnetic soft membrane actuator as an active microchannel component that can dynamically modulate its shape under magnetic fields and therefore the microfluidic environment. The membrane is magnetically programmed using a template-based approach, which allows it to be controllably deformed in the form of a sinusoid under the influence of an external magnetic field. The membrane's integration into a microchannel converts a passive channel wall into a dynamically changeable interface, allowing active fluid manipulation and enhancing micromixing in laminar flow conditions. The proposed approach establishes a versatile platform for wirelessly controlled deformable interfaces in next-generation microfluidic and lab-on-chip systems.

Imitation learning has emerged as a powerful paradigm for learning visuomotor policies, but its generalisation and stability are limited by the scale and quality of demonstration data needed. A promising direction is to leverage more abundant but heterogeneous data sources, which differ in action space and often lack action labels altogether. Existing co-training approaches that combine heterogeneous data sources rely on heuristic and hand-engineered alignment techniques. In contrast, we argue that action representations should be grounded in prediction: actions that produce the same effect on the environment should share the same representation, regardless of their sources. To this end, we instantiate this principle by using a grounded latent-action world model (GLAM), a pair of generative models with a shared latent action space across data sources that is grounded by predicting future observations consistently across sources. This latent action space is used to train downstream behavioural cloning (BC) policies which map observations to latent actions and decode them back to robot actions, providing a paradigm for learning from heterogeneous data. Empirically, we demonstrate that GLAM successfully learns an aligned latent action space that facilitates action transfer across data sources with and without action labels. Across five manipulation tasks in simulation and in the real world, GLAM-aligned policies significantly outperform BC baselines and prior latent-action methods, achieving an average of +48% improvement in task success rate with the same data-scarce setting. Videos and code are available at https://viccccciv.github.io/glam/.