The authors propose Compounding Physical RSI, a self-evolving policy framework that uses automated capability-gap assessment to locate the task stage where a failure occurs, refines policies with behavior-regularized offline reinforcement learning, and builds candidates in weight space to preserve existing skills.
The paper studies using learned state-action-effect relations as offline contracts to detect action-interface mismatches that are compliant in shape and range but semantically wrong (axis permutation, sign inversion, rotation rescaling, gripper polarity reversal), and to rank the fields likely to be at fault.
The research team learned an adiabatic spectral submanifold (aSSM) model of the full body of a trimmed-helicoid soft trunk robot from data and used it in model predictive control (MPC), completing a range of shape control tasks on hardware. In closed-loop dynamic shape regulation, the method outperformed existing linear baselines by a factor of 10; in human-in-the-loop interactive demos, it completed section-by-section shaping and balanced a tennis ball for more than 25 seconds, whereas the linear baselines failed to complete the task. A model-free reinforcement learning policy trained on hardware for 12 hours failed on the same balancing task.
The authors propose soccerbot, a hierarchical reinforcement learning framework that distills a structured latent motion prior from large-scale human soccer motion capture data, constraining policy learning within a dynamically feasible motion manifold. A single skill-conditioned policy thereby learns moving, dribbling, trapping, and kicking in a unified way, while maintaining human-like motion and fluid transitions during skill switching. After robustness-oriented simulation design, the framework transfers zero-shot to the Unitree G1 humanoid robot, completing the full soccer skill set in real-world environments.
The authors propose H^2MARL, a hierarchical heterogeneous multi-agent reinforcement learning framework for humanoid whole-body control. The method performs policy factorization with physical heterogeneity awareness at the joint degree of freedom (JDoF) level, conditioning each torque-actuated joint on descriptors of its own actuation, kinematic, and dynamic properties, and uses closed-loop policy-executor co-optimization to jointly adapt the high-level pose planner and low-level joint executor with shared interaction trajectories.
The study proposes a failure-boundary profile evaluation method in which each scenario family varies only one measurable factor (such as corridor width or the time gap in oncoming traffic), and each variant is labeled from measurement as one where the agent should proceed or one where it should stop. Along this axis, two curves are estimated for each agent: commitment (the probability of attempting the passage) and competence (the probability of passing without contact). The method locates where each curve crosses 0.5 and gives intervals that account for the simulator's run-to-run variance, while also using certified labels to derive a decision accuracy metric.
TwinBridge proposes a digital twin-driven framework that provides policy initialization and dense visual rewards for real-world online reinforcement learning. The method builds digital twins by combining 3D Gaussian splatting with physics simulation; simulator states provide task events and geometric progress, which are used for pretraining with simulated demonstrations and for reinforcement learning, then distilled into a visual reward model. Combined with execution history, this computes dense rewards and supports human-in-the-loop policy optimization.
The study proposes a unified planning and control framework for mobile continuum robots, combining a global layer consisting of A* search and minimum-jerk planning with a local layer consisting of a geometry-consistent inverse kinematics solver and Riemannian optimization, to coordinate the mobile base and continuum deformation for obstacle avoidance. The authors compare it with multiple baseline planners in simulated pipe-traversal tasks and conduct real-world experiments on navigation among dense obstacles and manipulation in confined spaces. Results show improvements in motion consistency, feasibility, and adaptability.
The authors propose a learning-from-demonstrations method that jointly learns a dynamics model and object-specific latent representations of physical properties, and conditions an imitation policy on online estimates of these representations, enabling the robot to adjust manipulation actions such as regrasping and contact locations based on physical-property information acquired through interaction. During execution, the dynamics model and policy weights remain fixed, and only the latent representation is updated based on the discrepancy between predicted and actual observations. In real-world and simulation experiments with objects having different centers of mass, online adaptation improved task success rates, but in simulation the success rate remained about 30% at center-of-mass positions not covered by the demonstrations, indicating limited adaptation to unseen physical properties.
This study examines the visual stabilization performance of boundary-layer sliding mode control (SMC) relative to causal baselines such as LPF and PID under plant-binding conditions, where disturbance velocity approaches or exceeds the actuator slew limit and the per-step actuation budget is constrained.
The paper proposes learning to holistically predict future hand-object states from human-object interaction (HOI) datasets, using the predicted future object states as a general-purpose reward term for reinforcement learning, reducing reliance on task-specific reward engineering. The authors experiment on three manipulation tasks in simulation and the real world, reporting better success rates and generalization to novel objects than existing SOTA methods, and validate cross-embodiment compatibility by deploying skills to different robot hands. The work has been accepted as an oral paper at AAAI 2026.
This paper proposes a lightweight reflection-aware front-end for feature-based stereo visual-inertial odometry (VIO) that estimates ground height in real time and rejects only feature candidates whose inferred 3D positions lie below the estimated ground. The method uses a ground-region mask to restrict the validation scope, constructs frame-level ground-height measurements from stereo height uncertainty and pose uncertainty, stabilizes the estimate with a scalar Kalman filter with NIS gating, and performs selective rejection with an uncertainty-aware safety margin.
The paper proposes scoring each transition in the replay buffer with a statistical leverage score and drawing mini-batches from a mixture of this score and uniform sampling, to alleviate the under-sampling of rare events in robot offline experience. When actions are stored as one-hot encodings, the method instead uses the feature column with the largest kurtosis, and it drops the importance-weight correction of prioritized replay, directly optimizing a rebalanced objective; for rare cases related to the target, it applies the same tail rule to short trajectory windows.
The paper proposes Future-Window Inverse Kinematics (FWIK), an online inverse kinematics selection method that considers several future waypoints, connects multiple IK candidates in a layered joint-space graph, ranks them with a motion-time estimate, and then uses a learned model to remove transition paths prone to self-collision.
The author examines the applicability of RRT* to triangle-mesh terrain, noting that the original RRT* cannot run directly on a mesh because the surface has zero volume, so uniform sampling never hits it. The proposed Mesh RRT* solved every run on terrain with slopes up to 75 degrees, but its cost function is biased: after 30,000 iterations the reported cost is on average up to 2.7% below the exact geodesic, and the returned path is on average 0.1-0.5% longer.
The study proposes RACE, a robot policy checkpoint scoring metric that requires no real-world rollout. It uses a fixed set of held-out successful trajectories to calibrate action deviations against state-conditional thresholds, then applies a running product to reduce the weight of later reference predictions after large deviations. In simulation, RACE outperformed baselines across different validation sources, while some baselines degraded noticeably when the validation set changed; on a real robot, its rankings were consistent with deployment success rates across the evaluated policy backbones and tasks of varying difficulty.
ZipDrift proposes using a one-step drifting generator to learn cost corrections for short-horizon MPC, trained only on cost evaluations of candidate corrections without backpropagating through the optimizer.
The authors propose a multi-robot closed-loop adaptive sampling framework in which a single learned generative model performs forecasting, uncertainty estimation, planning, and measurement assimilation simultaneously. The model generates ensemble forecasts in latent space and decodes them into physical ocean states, uses ensemble statistics to construct A- or D-optimal sampling objectives, selects sampling locations within the robot-reachable range constrained by predicted currents, and feeds measurements back into the latent state through observation-conditioned diffusion sampling.
OUCH is a wrist-worn pneumatic haptic stimulation device that uses stacked pneumatic pouch actuators, a rigid indenter, and a compliant spacer designed to buckle at a set load, so that the indenter contacts the skin only during actuation. At 200 kPa pouch inflation pressure, its indenter's contact pressure is estimated to reach 18,500 kPa, more than 25 times that of a commercially available handheld pressure algometer tip, while requiring less total force. In preliminary perceptual testing, the force required for this indenter to elicit discomfort was about 89% lower than that of the pressure algometer, providing a basis for controllable, repeatable noxious stimulation at the wrist.
The authors propose OdorSpaces, a physics-simulation-based platform and benchmark for visual-olfactory embodied navigation. It integrates 3D odor transport and dispersion simulated by Fire Dynamics Simulator (FDS) into Habitat-Sim, enabling agents to navigate realistic indoor environments using both visual and olfactory observations.
The paper proposes the PHIL framework, which lets operators preview a policy's predicted future actions through an XR interface, take over to correct before undesirable actions occur, and pair rejected predictions with corrective actions for preference post-training. On four real manipulation tasks, with five correction episodes per condition, mean success rates rose from 44.0% to 70.3% for the diffusion policy and from 57.8% to 76.9% for X-VLA, compared with the HG-DAgger baseline; policies trained with proactive corrections showed fewer unnecessary recovery behaviors in successful episodes.
This paper proposes a contact-implicit whole-body optimization framework for humanoid robots to generate dynamic motion trajectories on parameterizable complex terrains such as steps and ramps, without requiring a predefined contact sequence. The method approximates ground geometry with smooth surfaces and extends rigid-body interaction from planar surfaces to arbitrary smooth surfaces through new contact constraints, using nonlinear programming to plan motions across various terrain topologies. The authors conducted preliminary feasibility tests on non-coplanar terrains in simulation and on the real robot ErgoCub.
When running reinforcement learning with prior data, a behavior constraint is important for trading off exploration against staying close to known good parts of the data. But choosing this constraint is delicate—we neither want to merely match all prior data nor to deviate too far. In our new paper, we propose a new constraint that combines a support constraint with a continuous penalty to achieve excellent offline-to-online reinforcement learning performance.
SkillMimic uses multiplicative Gaussian kernel rewards to turn pose, interaction, and contact into multiplicative joint conditions, learning 19 types of interactive skills from basketball demonstrations. Its hierarchical version uses discrete skill selection to drive a frozen low-level policy and reproduces four types of tasks on a new-generation simulation platform.
PyRUA-DataFarm proposes having a coding agent rebuild tasks in simulation, with a video model rendering them into real-ready data, replacing teleoperation for collecting robot demonstrations. On five real-robot tasks, policies trained only on this data achieved 93/100 successes, versus 84/100 for policies trained on teleoperated data; under an unseen tablecloth and colored lighting, the former kept 87 successes while the latter fell to 46.
Why it matters: The paper presents a complete pipeline and cost comparison for using a coding agent to replace teleoperation in collecting demonstrations, which can be used to assess the room for replacing this data-collection route.
The team led by Wu Xinyu at the Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, has developed a wearable underwater exoskeleton. Its representative system, PEAKED, is a soft underwater hip exoskeleton that assists hip flexion and extension, supporting flutter kicking, breaststroke kicking, and underwater walking. In prior human experiments, PEAKED reduced per-minute breathing volume by 29.77% during flutter kicking, 25.70% during breaststroke kicking, and 18.35% during underwater walking.
At Rogers Dry Lake in Edwards, California, NASA completed simulated lunar descent and landing maneuvers using an Alta-X drone carrying the SPLICE guidance, navigation, and control payload. The SPLICE payload uses its guidance system to map terrain and assess landing zones, redirecting to a safer location when it determines the target area is unsafe, and generates high-resolution digital elevation maps from light detection and sensor data.
Google and Mercor released the Physical-Engineering's Last Exam (PLE) benchmark, which evaluates engineering agents on 85 mechanical, electrical, civil, and structural engineering tasks.
D-Robotics, together with the University of Tokyo and other institutions, released the RawVLA technical approach, which adds a learnable neural ISP module before the VLA policy while keeping the VLA model frozen, using action error to train the imaging front end in reverse. On real robots, four dual-arm tasks were tested with π0.5; the average success rate in low light rose from 0% with the default ISP to 66.5%, versus 75.0% in normal light, where the default ISP achieved 76.5%.
Why it matters: The approach brings the imaging front end into the learnable pipeline and provides comparative data under low light and normal light, offering a reference for vision pipeline design.
A team at Tohoku University has developed a mechanical reflex module that requires no electronic sensors or circuits. Integrated into a musculoskeletal robot, it reproduced the knee-jerk reflex, and combined with high-level control that adjusts activation thresholds according to movement phase, it achieved bipedal walking. The module integrates sensing, computation, and actuation within its physical structure: it mechanically detects when an artificial muscle is stretched beyond a threshold and contracts automatically, inspired by human muscle spindles.
BSC-Nav is an arXiv paper (2610.02784) aimed at indoor object navigation, long-instruction following, active embodied question answering, and mobile manipulation, addressing the problem that multimodal large models can understand the current view but struggle to support spatial cognition. This analysis article from Guyuehome says it moves embodied intelligence from reacting to a single glance toward retrievable spatial cognition, but currently only the abstract is provided; method details and experimental results are not elaborated.
EvolvingNav proposes a dynamic search framework to address the problem of robots tracking items that may be moved in long-term environments. The framework fuses timestamped 3D observation history and distinguishes three cases: the target remains at its most recent location, the target has moved to a known location, and the target has entered an unknown area.
The article reviews three paths for embedding physical priors into robot learning: data and representation, model architecture, and training loss, and emphasizes that where a prior is placed determines its constraining effect. Using grasping a soft object as an example, it explains the layered roles of kinematics, dynamics, contact, and material properties in the task, while noting that physical priors are not a panacea and need to match the task.
RPV-SemNav proposes a zero-shot indoor navigation method that requires no prior map. It maps target and observed objects into probability vectors over nine room categories (RPV) and uses spatial co-occurrence semantics to guide exploration. The system combines open-vocabulary detection, CLIP semantic alignment, and free-space propagation, enabling the robot to first understand the room and then decide where to look for the target.
SSTG-Nav proposes a reusable object navigation framework for long-term deployment scenarios, focused on addressing the mismatch between observation poses and docking poses. The framework uses a single task-agnostic RGB-D mapping pass, employing a vision-language model (VLM) for detection and depth back-projection to generate reachable targets. The material provides only an abstract and does not give full experiments or results.
Super Odometry 2.0 introduces a four-layer adaptive mechanism for motion estimation when cameras and LiDAR degrade in scenarios such as darkness, repetitive corridors, and dense smoke. The system filters reliable observations at the feature layer, identifies weakly constrained directions at the direction layer, and dynamically adjusts the factor graph fusion strategy at the engine layer.
Flapping Flyer reproduces a 77 mg flapping-wing aerial vehicle in MuJoCo, modeling it with a three-term blade-element aerodynamic force composition, and achieves hovering with LQG closed-loop control, with altitude fluctuation of 0.12 mm. The material also mentions wing wash, but the main text does not elaborate on it.
The study proposes Route Anything, a unified framework for studying model and policy routing in general-purpose AI systems, and establishes an optimal routing theory under quality-cost trade-offs. The authors construct an omni-modal trajectory matrix covering 25 benchmarks across language, image, video, speech, coding and workplace agents, and robotic manipulation, with 452 candidate configurations and 374,401 scoring runs.
The research team proposes PrimeSim, an action-conditioned multiview world model simulator, midtrained on robot demonstration data covering 22 platforms and 30,110 hours, and has released the model weights, inference code, and training code.
Why it matters: This work presents a world model simulator trained on demonstrations spanning 22 platforms and 30,000 hours, and releases the weights and training code, facilitating comparison of its long-horizon stability and action-following capability.
DreamFree proposes a world action model (WAM) that does not require generating future videos. Through representation distillation, it turns a pretrained video diffusion model into a predictive visual encoder that predicts the frozen video model's representation using only the current observation; an action module then extracts action-relevant information.