Rongxuan Zhou

Contact-rich robotic manipulation has long relied on vision-based policies, but at the contact phase vision is often insufficient: the end-effector occludes the workpiece, and under commodity sensing the last millimeter is hard to close from vision alone. Force/torque sensing complements vision at exactly these moments, yet it is often unavailable at deployment: training datasets come from sensor-equipped platforms while target robots mount no force/torque sensor, and wrench estimates drift out of calibration. Existing force-conditioned policies treat force as an always-available input and do not address a missing sensor at test time. A natural remedy is to add an encoded wrench into the latent state of a vision world model, so that the model uses force when present and falls back to vision otherwise. In practice this backfires: on our force-augmented PushT benchmark (PushT-FT), planning success collapses from 41% to 15% once force is missing at test time. The collapse persists across fusion operators and even when missing force dominates the training data. What matters is the amplitude at which force is written into the shared latent. To address this, we propose ForceLeWM, a JEPA world model that writes force into the latent through a conservative per-dimension gate at a small learned amplitude, trained with a warmup alignment loss for the from-scratch setting. ForceLeWM restores planning to 35%, recovering about 75% of the lost planning success across 3 training seeds. A train-consistent normalization audit shows the perception benefit of force is task-specific: clear on FORGE peg insertion but not across other task families. We further deploy one frozen simulation-trained checkpoint end-to-end on two real Franka assembly tasks, gear meshing and peg insertion, with force supplied by a proprioceptive wrench observer estimated from joint torques rather than a dedicated force/torque sensor. When test-time force availability is uncertain, fusion should be conservative by design: admit force at a small learned amplitude when present, and degrade gracefully rather than collapse when absent.

In latent world models of the joint-embedding family, the planner learns about actions only through the model, and the model compresses each action block through a linear bottleneck. The compression is uneven over action directions, and along some directions it is total. We call the non-uniformity action-compression anisotropy. Existing treatments of action insensitivity modify the training objective and leave released, frozen models undiagnosed. This paper introduces an external protocol that diagnoses them. It reads per-direction sensitivity from end-to-end Jacobians, compares it with the response the environment exhibits along the same directions, and measures how much control the planner exerts along each. On four released checkpoints the protocol separates two kinds of compression. Compression learned by training is faithful, with sensitivity matching environment response at correlation 0.87 to 0.97. Compression imposed by the architecture is a blind spot, and one checkpoint is exactly insensitive to fifteen directions carrying about 8% of the environment's action response. The diagnosis supports a training-free intervention. Projecting executed actions onto the model's visible subspace raises cube manipulation success from 68.5% to 75.0%. The gain has a stated validity boundary. The projection helps only where the planner has no measured control over the deleted directions, and it costs up to 18.0 points where weak control survives. The state channel offers no such option. Projecting the planning cost onto the reachable subspace loses 18.7 to 25.3 points, and no tested rank recovers its baseline. A restriction is safe when the discarded component carries no planning signal, and that can be measured with everything frozen.

Vision-Language-Action (VLA) models have advanced robotic navigation by unifying perception and planning, yet most systems operate in an open-loop manner, generating plans without awareness of locomotion feasibility or the ability to adapt when execution becomes unsafe. This disconnect leads to cascading failures on challenging terrains and surfaces where pre-planned trajectories become infeasible. We propose Thinking on Its Feet (TOIF), a three-tier hierarchical control framework that closes the loop between low-level proprioceptive sensing and high-level semantic path planning for terrain-aware humanoid robot navigation. At the low level, an RL locomotion policy is augmented with a multi-detector anomaly system that evaluates terrain-specific feasibility and safety constraints using compact proprioceptive features. Detected anomalies are accumulated through leaky integrators and translated into structured natural-language feedback, enabling the high-level planner to replan over a semantic scene graph. At the mid level, a VLA model generates velocity commands that bridge high-level waypoints and low-level execution. We further introduce a risk-aware velocity modulation mechanism that proactively reduces speed based on real-time anomaly confidence. We validate TOIF in Isaac Lab simulation and on a physical humanoid robot across a range of challenging terrains. Results demonstrate that TOIF substantially improves navigation success rate and safety over open-loop baselines, and that the proprioceptive anomaly detectors transfer effectively to real-world deployment.

Unfolding a crumpled garment from an overhead view is ambiguous: a flat cloth and a stack of hidden folds can share the same silhouette. This paper studies a transparent-table dual-arm setup in which the robot observes both a top segmentation mask and a bottom contact mask. The main finding is that the bottom view is most useful when converted into fold-contact affordances rather than used as a generic second image. We derive fold residual, contact overlap, fold-contact frontier, and silhouette-edge maps from the dual masks, use them in an attention encoder, and further convert the frontier into a residual action prior. To reduce hand tuning, we parameterize the prior's cell scorer as an online linear, contextual-bandit-style update trained from reward, self-supervised unfolding progress, or their mixture. Across the completed logs, raw dual-view pixels improve long-horizon flatness over a top-only policy, affordance channels reduce folded area by 82.2% relative to raw dual masks in a fast screen, and a reward-bandit bottom-view scorer gives the best 21-run residual-prior ablation on within-panel geometry score. The results support a precise conclusion: bottom view helps cloth unfolding by exposing where upper folded layers meet grounded support, and reward-trained scoring is the most effective tested way to turn that information into manipulation targets.

Diffusion policies have become the dominant paradigm for learned robotic manipulation, yet they offer no intrinsic mechanism to enforce workspace safety constraints. TACS is a training-free method that injects barrier-function guidance into the DDIM denoising loop of action-space diffusion policies, applying repulsive guidance on the Tweedie-predicted end-effector positions; subsequent denoising steps coherently adapt rotation and gripper dimensions, achieving effective obstacle avoidance at 10× smaller guidance scales than post-hoc potential-field projection. On three robosuite manipulation tasks, TACS Pareto-dominates post-hoc filtering, with a statistically significant win-win on Lift (violations 6.6% → 0.32%, 20× reduction; p=0.015 across 5 seeds × 250 episodes) and zero measurable inference overhead versus the unmodified diffusion policy.

Odometry estimation remains a critical challenge for wheeled robots, as reducing its drift directly mitigates dependency on external localization systems. This paper proposes a distributed odometry framework for steerable wheels, named ICF-DO, which is applicable to both Steerable Wheeled Mobile Robots (SWMRs) and cooperative multi-single-wheel robot systems. The proposed method features low computational complexity and reduced drift, while demonstrating strong robustness in communication-restricted scenarios. Additionally, singularity can be processed in a distributed manner in the proposed framework. Experimental validation on a real physical SWMR platform demonstrates the effectiveness and practicality of the proposed method.

The vital component of the high-pressure pump is the high-pressure cylinder, which undergoes pulsating cyclic loads during operation. This exposure can lead to fatigue cracking and limit the pump’s overall performance. To address this issue, a swage autofrettage treatment method for the high-pressure cylinder is proposed based on the principles of autofrettage technology. The structure parameters of the mandrel are designed to optimize its role in the treatment process. Subsequently, a swage autofrettage simulation model is established, considering the Bauschinger effect to analyze the distribution of residual stress about the mandrel interference percentage. The results demonstrate that the swage autofrettage treatment improves the stress distribution in the high-pressure cylinder, causing the inner wall surface to undergo a circumferential residual compressive stress. Within a specific range, the Bauschinger effect becomes more significant as the mandrel interference percentage increases. This study provides insights for enhancing the durability and performance of high-pressure cylinders while adhering to mechanical engineering standards.