A flow-matching policy deployed in the field must often satisfy a constraint that was absent from its demonstrations, such as avoiding an obstacle it has never seen. Retraining for every new constraint is impractical, so the constraint must be enforced at inference time. Enforcing it must not discard what the policy learned: an action the policy would rarely sample may avoid the obstacle yet fail the task. We therefore want actions that are both safe, satisfying the constraint, and plausible, remaining likely under the policy. Gradient guidance, the most common approach, steers noisy states, so it must evaluate its cost on the policy's running estimate of the clean action; when the policy is uncertain between passing left or right of an obstacle, this estimate is their average and can pass straight through it. We instead build on rejection and projection. Rejection redraws actions until one is safe: every output is plausible, but the wait is intractable when safe actions are rare. Projection moves an action to the nearest safe one: it is efficient, but the result need not be plausible. Our key insight is that flow-matching sampling lets us combine the two, keeping projection's efficiency and rejection's plausibility. CS-Pi, our draw-project-renoise sampler, projects each draw to the nearest safe action, adds noise back scaled to how far the projection moved it, and re-runs the unmodified policy until a draw is safe. Replacing only the projection step extends CS-Pi to continuous costs. Across a 2D diagnostic, a 7-DoF shelf pick-and-place, an orientation preference, and a zero-shot Franka deployment, CS-Pi reaches 81% safe success on the shelf and 55% on hardware, against 55% and 40% for projection, while staying plausible.
https://arxiv.org/abs/2611.08654
As robots are deployed in the real world, they will inevitably make mistakes. Human interventions provide valuable feedback for improving their behavior, enabling them to adapt online to new scenarios. However, studying human-in-the-loop robot learning is difficult: real-world experiments are slow and expensive, so most algorithm development relies on simulated human models. We show that common human modeling assumptions do not match how real people intervene in practice. Through analyzing intervention data from human participants, we find that humans tend to intervene when agent progress stalls, and that their corrective actions are often suboptimal. Motivated by these findings, we introduce a progress-based model of human intervention that achieves higher rank correlation with real-user algorithm rankings, enabling realistic algorithm benchmarking and development in simulation. Building on these insights, we propose Q-OIL, a method that adds optimistic bonuses to intervention transitions in reinforcement learning and softly guides exploration toward human corrections. We find that decoupled critics are critical for stability because they prevent the bonuses from backing up to pre-intervention states. As a result, Q-OIL learns more efficiently than competitive baselines, achieving 10-50% higher success in simulation and on real robots, and improving pretrained VLAs on two bimanual manipulation tasks. As we ground both evaluation and algorithms in realistic human behavior, this work provides a practical path toward scalable robot learning from interventions.
https://arxiv.org/abs/2611.08576
Fine-grained, bimanual dexterous manipulation remains a foundational challenge in robotics. Embodiment gaps hinder accurate kinematic mapping, and visual tracking alone does not regulate contact forces. To address these limitations, we propose a tactile-driven adaptation strategy designed to enable fine-grained manipulation on top of teleoperation pipelines. Instantiated within our bimanual dexterous framework, DexTeleop-0, this strategy introduces a real-time optimization loop that adds robot-side contact corrections to human tracking commands using fingertip feedback. By estimating contact positions and forces, the system dynamically computes localized corrections using a contact-stiffness approximation and the kinematic Jacobian, with per-contact vector force tracking, soft predicted-contact-moment regularization, and bounded hand-joint residuals. We evaluate this tactile-driven adaptation strategy across both simulated environments and real-world hardware. Controlled simulation replay yields final-stage success rates of 91% for Ball Assembly and 62% for Stir in Cup, versus 88% and 49% for force tracking alone. Fixed-input replay on two hand types, seven-operator teleoperation, and external-force traces further assess contact assistance. The open-sourced project page can be found at this https URL.
https://arxiv.org/abs/2606.23431
We present cuRoboV2-Fast, a GPU-accelerated planner that reduces both planning latency and execution time for collision-free, payload-aware, torque-limited motions. B-spline trajectory optimization reduces the required iterations, while accelerated robot computations and differentiable inverse dynamics reduce refinement cost and accommodate changing payloads. TimeOpt optimizes traversal speed along the fixed path under sampled velocity, acceleration, jerk, and torque constraints, exploiting local spline support to solve banded Newton systems efficiently on the GPU. With a 3,kg payload, cuRoboV2-Fast achieves 4.4 ms mean planning time on RTX 5090, 14.7x faster than cuRoboV2. Our method on Jetson AGX Orin plans faster than cuRoboV1 + TOPPRA on RTX 5090. A stereo-camera-based deployment on YAM achieves 5.93 s mean completed pick-and-drop cycles, 34.6% shorter than cuRoboV2. Code (Apache 2.0): this http URL , Project Page: this http URL .
https://arxiv.org/abs/2611.06913
The introduction of generalist robot policies has enabled a single, large model to learn to perform a variety of tasks from a vast pre-training corpus. While these models often exhibit powerful zero-shot generalization capabilities, they typically require an adaptation phase to achieve sufficient performance in deployment. Iterated self-improvement via reinforcement learning (RL) has demonstrated promise in enabling this adaptation, yet such approaches have been largely explored in single-task settings - where the goal is to simply improve on a narrow deployment task - which lies in stark contrast to the multi-task capabilities of the base policy. In this work we ask whether generalist robot policies are capable of multi-task, generalist improvement in deployment - targeting many diverse tasks simultaneously - and what algorithmic recipe enables this. To answer this question, we first develop a set of structured evaluations to test generalist improvement capabilities, and systematically evaluate algorithmic and practical design choices regarding value estimation, policy extraction and resource allocation. We show that not only can generalist policies be improved online across diverse tasks, but that the right recipe can enable generalization and increase performance on tasks not considered during online training. We characterize when this generalization occurs and show that both the training task distribution and the use of value-based RL are critical. To the best of our knowledge, this is among the first evidence for online generalist improvement for robot foundation models.
https://arxiv.org/abs/2611.06905
Conventional robot arms can be extremely heavy compared with the everyday objects they manipulate, creating a significant inertia mismatch at contact. Even a small accidental bump from a heavy arm can knock a light object out of place. Modern learning methods have nevertheless demonstrated a wide range of successful manipulation behaviors on these platforms, suggesting that even extreme inertia mismatch can be overcome with sufficient data. However, it remains unclear whether high inertia imposes a hidden cost on learning, potentially worsening scaling with data or lowering the performance ceiling of modern robot learning methods. We investigate this question through controlled simulation experiments that vary robot and object masses. We find that reinforcement learning succeeds across a broad range of mass combinations, while behavior cloning learns more reliable policies when a lighter robot manipulates a heavier object. Heavier robots require more demonstrations to attain the same performance as lighter robots. Finally, because new hardware is rarely an option, we explore ways to lower an existing arm's inertia at contact with existing hardware, either by choosing a lower-inertia workspace before collecting data or by steering the demonstrator toward lower-inertia configurations during the task. In both cases, the lower-inertia demonstrations are imitated more reliably. Project website: this https URL
https://arxiv.org/abs/2611.03501
For robots to operate reliably in commercial and industrial applications, they must persistently perform specific tasks with known variations in object shape and pose. We call this Variational Automation (VA) to differentiate these settings from fixed automation. Inspired by recent papers on Agentic Robotics, where pretrained LLM and VLM coding agents combine interpretable robot programming with the open-world adaptability of model free policies, we propose a general set of agentic robotics benchmark design criteria that includes accessibility, repeatability, diversity, relevance to agentic robotics and practical applications, physical realizability, resistance to benchmark saturation, and computational efficiency. We instantiate these principles in six manipulation task classes spanning industrial, medical, and infrastructure applications: tool sorting, tool hanging, gear meshing, syringe packing, cable insertion, and cable routing. We introduce the Variational Automation Benchmark for Agentic Robotics (VABAR), an open, easily replicable, simulation-based manipulation benchmark to systematically evaluating agentic robotics systems. Built fully on open-source tools, VABAR features detailed implementations using the Newton physics engine, standardized task variations, difficulty tiers, success metrics, and a library of 33 generalizable skills with human-authored programmatic reference solutions. We provide initial baseline evaluations using an open agentic manipulation harness, Graph-as-Policy (GaP), and Isaac Lab Arena.
https://arxiv.org/abs/2611.06881
Code-as-policy agents often use execution feedback to retry failed programs. Additional sampling and task guidance can also improve success, leaving the contribution of failure-specific feedback unclear. We study this question in simulated robot manipulation. A fixed lookup selects feedback templates using failure labels and the API variant. A frozen code model generates a fresh program after each scene reset. On the same 252 initial failures, Qwen2.5-Coder-7B recovers 85.3% within five retries, compared with 68.3% for resampling with the same retry cap. The advantage persists under fixed total-token budgets. Feedback from other episodes and fixed task guidance yield similar recovery point estimates, with no detected benefit from matching the coarse failure label. In a separate study, a fixed guided prompt used from the first attempt achieves an observed overall success rate similar to that of template-based feedback. Template-based feedback also lowers recovery on a nut-insertion task requiring a rim grasp, consistent with an incompatible center-grasp instruction. Evaluations of robot self-repair should include resampling, mismatched feedback, and concise, task-compatible static guidance as baselines.
https://arxiv.org/abs/2611.06863
Moving robot swarms from simulation into the physical world can be difficult because small discrepancies in hardware, timing and communication can cascade across interacting robots and alter group behaviour. This challenge is known as the reality gap. When swarms operate in human environments, a second challenge emerges: a perceptual gap between what the swarm physically does and what human onlookers actually perceive. This study examines the same decentralised coordination code deployed across three robotic swarms-tabletop, aerial and human-scale mobile manipulators-during live artistic performances. We find that while the underlying algorithms executed reliably on all three platforms, physical embodiment drastically altered visible motion-distorting execution timescales by more than a factor of three and reshaping trajectory geometry. More importantly, audience surveys revealed that human interpretation does not mirror technical success: how observers attribute meaning (perceived purpose) and how they react emotionally (comfort versus threat) depend heavily on physical form. Meaning attribution and emotional responses did not vary together. Our results demonstrate that successfully transferring swarm software to physical hardware is insufficient for real-world deployment. Addressing the reality gap does not guarantee that the resulting collective motion will be readily interpreted by human observers. Swarms destined for human environments must be engineered not only for decentralised stability but also for the expressive physical dynamics through which humans interpret collective motion.
https://arxiv.org/abs/2611.05244
Hamilton-Jacobi (HJ) Reachability identifies the unsafe states of a dynamical system from which it can be forced into failure under worst-case disturbance. However, classical grid-based solvers scale exponentially with state dimension. DeepReach improves scalability by replacing the grid with a neural value function trained on the HJ-Isaacs (HJI) variational inequality (VI), but penalizes both sides of the residual equally. This allows the learned value to lie above the true one and unsafe states to be misclassified as safe. We propose Conservative DeepReach (ConDR) which aims to learn a subsolution of the HJI VI such that the learned unsafe set contains the true one. ConDR enforces the obstacle branch and terminal condition by construction, pushes the HJI residual to a safety margin $\epsilon$ below zero with a one-sided loss, and prevents the degenerate solution of classifying everything as unsafe. We prove that the exact solution of the margin-shifted VI that ConDR trains toward is a subsolution of the original VI so its unsafe set contains the true one and its value lies within $\epsilon\tau$ of the true value function. We compare against DeepReach and MADR in the established systems of Air3D, Air6D, a 9D multi-vehicle system, and a 13D quadrotor. ConDR reports the lowest rate of unsafe state misclassified and the lowest closed-loop collision rates with at most a small increase in intervention. On Air3D and Air6D, ConDR tolerates model mismatch that scales with $\epsilon$. At matched false-negative rates, ConDR's unsafe sets are smaller than the baselines showing the safety is not obtained by trivial over-approximation.
https://arxiv.org/abs/2611.06802
Coding agents powered by models such as GPT-6-Astra connect spatial reasoning with executable robot behavior, raising the question of how such agents can produce lasting improvements in robot capabilities from physical interaction. In this work, we study embodied recursive self-improvement (RSI) through real-world exploration, code refinement, and deployment using PyRUA-REAL, a framework in which coding agents retain interaction experience as persistent memory and reusable executable skills that inform subsequent exploration and improvement. Our empirical study places physical manipulation at the center of the improvement process, examining how agents collaborate, what kinds of execution feedback support effective code refinement, and how task specifications and previously acquired skills influence robustness and task transfer. We further investigate the efficiency and capability of the resulting programs when deployed without involving an online agent for decision-making. Across eight real-world tasks, PyRUA-REAL achieves an average success rate of 91.1%. In complementary evaluation on 42 tasks from the RoboDojo benchmark, it achieves an average success rate of 84.4%. This study provides empirical guidance for turning physical exploration into reusable robot skills that support continued improvement and robust, efficient deployment.
https://arxiv.org/abs/2611.06768
Vision-language-action (VLA) models map camera observations and language instructions to robot actions. However, their large backbones and repeated action generation can make inference take longer than the robot's real-time requirements for its next action. To reduce the load of VLA computations, researchers have proposed multiple optimization approaches, including quantization and pruning. These optimizations can be combined with other general optimizations such as better kernel compilation or using CUDA graphs instead of the eager mode. However, most of these optimization results are usually reported on a single GPU type, typically for a few VLA models, with no analysis of what the real bottlenecks are for a model on a particular hardware or how these optimizations would generalize beyond the reported results. In this paper, we provide a comprehensive measurement study of seven families of inference optimization methods, from CUDA graphs and compilation to token pruning and quantization, on seven state-of-the-art VLAs evaluated on four GPU families ranging from a datacenter H100 to an embedded Jetson Thor. We analyse how CPU launch overhead, memory bandwidth, and kernel efficiency jointly shape optimization gains, and how applying one method changes the benefit of another. From this analysis, we derive VLA-Triage, guidelines for stacking optimizations under latency budgets. The guidelines identify which stages can provide the required savings and which combinations need testing in the deployed execution mode.
https://arxiv.org/abs/2611.06710
Heterogeneous air-ground teams can use aerial sensing to address the limited range and viewpoint of ground-vehicle sensors, but deciding when the UGV needs assistance and what information the UAV should collect remains a challenge. This paper presents a planner-conditioned air-ground assistance method that evaluates how uncertain map information affects the UGV's current MPPI rollouts. When navigation remains degraded, we score the same rollouts under resolved map hypotheses, select the map layer with the largest useful improvement, request the corresponding aerial measurement, fuse the returned information, and replan. Unlike approaches driven by fixed support roles, task structure, or uncertainty alone, our method conditions assistance on whether resolving uncertainty improves predicted ground navigation. Across 180 simulation trials on three off-road routes, the method achieves 98.9% goal attainment while reducing mean UAV flight time by 87.7% relative to continuous greedy aerial mapping. Aggregated across the routes, each second of estimated UAV flight time saves 2.94 s of UGV navigation time. Field experiments spanning approximately 2 km of UGV operation demonstrate uncertainty resolution and successful navigation recovery.
https://arxiv.org/abs/2611.06702
Teleoperating a Remotely Operated Vehicle (ROV) in complex underwater environments presents unique challenges. Dynamic, three dimensional obstacles and unseen water currents pose real risks to expensive robotic hardware in environments in which recovery can be next to impossible. This paper explores the cognitive effects of haptic feedback to improve vehicle safety in such environments. An in-water study is analysed with a ROV exposed to controlled cross-currents, with environmental information presented via a soft and vibrating haptic touchpad. Results demonstrate that while encountering an unseen cross-current significantly increases an operator's physical and temporal demand, the introduction of haptic feedback mitigates these factors and improves self-assessed performance. Interestingly, the haptic condition also produced the highest levels of mental demand and operator frustration. Despite this increased mental workload, operators drove the ROV significantly faster when receiving haptic cues compared to the current-only condition, indicating a higher level of user confidence. Ultimately, these findings suggest that while haptic feedback may initially require more cognitive processing to interpret unseen hazards, it successfully enhances operator awareness and promotes more confident navigation in hazardous environments.
https://arxiv.org/abs/2611.06689
Vision-based tactile sensors capture multiple aspects of physical interaction within a single image stream, including contact appearance, geometry, indentation depth, and tangential deformation. However, existing visuo-tactile policies often directly process tactile observations as images, making it difficult to determine which physical cues contribute to policy learning. This paper presents a controlled real-robot study that compares physically interpretable tactile signals using the LeRobot platform, enhanced with our visuo-tactile perception system. From the same tactile stream, we derive tactile RGB images, normal/depth maps, and marker-flow fields, representing contact appearance, geometry, and tangential deformation, respectively. Using identical task demonstrations and the same policy backbone, we evaluate the individual and combined contributions of these tactile cues. We further compare tactile encoders and visuo-tactile fusion strategies. Evaluation across 720 real-robot trials yields four key findings: (1) tactile perception substantially improves the overall success rate from 57.8% to 78.9%, although its contribution varies across task stages; (2) marker flow is the most effective individual cue, achieving a success rate of 74.4% and playing a particularly important role in regulating established contact; (3) cue-factorized encoding consistently improves performance, highlighting that different physical cues require tailored processing techniques; and (4) although the effectiveness of tactile fusion is task-dependent, cross-attention offers a lightweight solution with performance comparable to that of pretrained tactile encoders.
https://arxiv.org/abs/2611.06673
Generalist robot policies are trained on static offline datasets, yet physical manipulation unfolds sequentially across stages. Because a weakness in a downstream stage remains unobservable until the policy passes earlier ones, each resolved bottleneck exposes the next hidden failure mode. This staged structure calls for Recursive Self-Improvement (RSI), in which a policy iteratively executes, diagnoses failures, and updates itself. Moreover, because task completion behaves like a product of stage-level pass rates, recursive gains can multiply across generations. In physical environments, three bottlenecks obstruct compounding: terminal execution feedback rarely localizes the stage that caused a failure, failed trajectories do not supply the corrective action, and updating shared parameters can degrade previously mastered skills. We propose Compounding Physical RSI, a self-evolving policy framework that turns execution failures into heritable capability. First, automated capability-gap assessment traces failed rollouts to the task stage where progression breaks down and formulates a structured improvement objective. Second, conditioned on this objective, our remedy stage curates validated, stage-relevant experience and refines the policy via behavior-regularized offline reinforcement learning, without requiring costly online rollouts at each optimization step. Third, to safeguard existing skills, our retention stage constructs weight-space candidates between the inherited reference policy and the refined policy, evaluates them across the task suite, and promotes the best candidate as the reference state for the next generation. Evaluated across all 42 tasks of the RoboDojo-Sim benchmark, CSU-AI-07 ranks first on the leaderboard with a three-seed mean Overall Score of 26.15 and Success Rate of 20.70%, outperforming the next-best entry by 5.92 points and 5.82 percentage points.
https://arxiv.org/abs/2611.06654
A robot action can satisfy shape and range constraints yet carry the wrong semantics: axes may be permuted, signs inverted, rotations rescaled, or gripper polarity reversed. We study learned state-action-effect relations as offline contracts for detecting these mismatches and ranking their likely fields. Clean held-out transitions calibrate alarms; a fixed inverse vocabulary supports closed-set localization. On ten LIBERO Spatial tasks comprising 432 source episodes, a capacity-controlled comparison at 2.5% empirical clean false-positive rate gives 46.1% recall for a ridge forward residual, 61.0% for a four-component joint Gaussian mixture, and 65.7% for the exact conditional of that same mixture. The joint score retains a 14.9 percentage-point advantage over ridge, while conditioning improves recall by a further 4.7 points. Joint scoring instead has higher localization: 95.8% versus 94.6%. These matched-rate comparisons are retrospective diagnostics, not deployment calibration guarantees. Earlier Object/Goal results broaden tasks within the LIBERO convention; DROID, all three HYDRA tasks, short temporal faults, small offsets, exact repair, and simulated mitigation fail their respective gates, while xArm contributes one task only. The evidence supports a bounded offline diagnostic and a detection-localization tradeoff, rather than joint-density superiority or cross-robot safety.
https://arxiv.org/abs/2611.06653
Wiping, sanding, and inspection robots press a sliding tool on a surface with controlled force. Controlling that force needs the surface normal, which is unknown on real-world, unmapped surfaces. Surface-normal estimators can help estimate the surface normal, but they either need a friction model or assume that the instantaneous tool velocity lies along the surface; however, walking violates this assumption. We propose SCRUB, a contact-based surface-frame estimator, which lets an armless quadruped slide a body-mounted tool along near-vertical surfaces in the absence of known, axis-aligned, or vision-predicted normals. SCRUB's predictive controller presses and slides the tool using a Kalman-filter estimate of the normal from two cues: 1) A force cue measures the estimate's tilt across the motion. If friction directly opposes the motion, the cue is zero exactly when that tilt is zero, for any friction coefficient. However, if the cue's sliding direction is wrong, friction appears as a false tilt. 2) A velocity cue measures tilt along the motion. Walking and pressing move the tool in and out. SCRUB therefore subtracts the commanded press rate from the tool velocity, averages the result over each gait cycle, and treats this mean as lying along the surface. On Unitree Go1 hardware at nominal speed and load, SCRUB completed every wall traverse and column orbit. No orbit completed with either cue removed, without gait averaging, or with a fixed normal.
https://arxiv.org/abs/2611.06649
Controlling the shape profile of soft trunk robots extends their reach into complex regions of the workspace and enhances their use in practical applications. Recent advances in soft robot design have enabled soft trunk robots to achieve greater flexibility in conforming to complex shapes. Specifically, trunk robots whose geometries are designed as helicoids have been shown to enable extreme shape-changing capabilities. Their compliance makes them highly nonlinear, which in turn makes shape control difficult. On similar tendon-driven trunk robots, model-based controllers based on the theory of adiabatic spectral submanifolds (aSSMs) have achieved high accuracy in closed-loop end-effector tracking. We extend this approach by learning an aSSM model of the full body of a trimmed-helicoid soft trunk robot from data and use this in a model predictive control (MPC) scheme to demonstrate a range of shape control tasks on hardware. On closed-loop dynamic shape regulation, our method outperforms the available linear baselines by a factor of 10. In human-in-the-loop interactive demos, our method completes section-by-section shaping and balances a tennis ball for over 25 seconds, while the linear baselines fall short. A model-free reinforcement learning policy trained for 12 hours on hardware fails on the same balancing task. These results establish aSSM-based MPC as a practical route to full-body shape control on real soft hardware.
https://arxiv.org/abs/2611.06625
Ground robots operating in temperate forests suffer from systematic LiDAR occlusions caused by tree trunks, creating persistent blind regions that degrade traversability estimation. In this paper, we propose FOSCo-Net, a self-supervised voxel completion framework that learns to reconstruct occluded geometry from a single scan using temporally accumulated observations as automatic supervision, and we evaluate cylindrical and Cartesian ego-centric grids with multiple decoder designs. FOSCo-Net consistently improves reconstruction quality and traversal safety, with a gated skip decoder achieving the most reliable performance and significantly reducing false obstacle predictions compared to baselines.
https://arxiv.org/abs/2611.06588