Effective planning requires representations that preserve task-relevant geometry and object identity while ignoring visual variation that does not alter the task. Existing evaluations of visual generalization often change several factors simultaneously, making the source of a failure difficult to identify. We introduce a controlled evaluation of object-goal navigation (ObjectNav), the task of reaching an instance of a named object category, in procedurally generated ProcTHOR houses. Training and evaluation houses share floor plans, object locations, goals, and start states, while visual factors are changed individually or in combination. Across model-free reinforcement learning (RL), model-based RL, and agents using frozen pretrained visual representations, our results show that appearance changes impair navigation, but their effects differ sharply by factor. Surface material changes account for most of the average change in success, whereas clutter and lighting have little effect. The magnitude of a global image statistics shift does not reliably predict performance: changing only the goal object's appearance can substantially impair an end-to-end agent despite a small image-level change. We found that a model based agent that plans over frozen features from the pretrained DINOv2 image encoder achieves the highest success (0.83) under the full shift, but remains sensitive to surface appearance. Our experiments reveal an important requirement for planning representations: they must separate invariant spatial task structure from appearance.
https://arxiv.org/abs/2611.19621
Reinforcement learning (RL) enables robots to acquire complex skills, but standard RL formulations do not guarantee that learned policies satisfy safety constraints, which limits their use in safety-critical settings. Existing safe RL methods with provable guarantees typically support only restricted classes of constraints and require unbiased estimators of the objective and constraint functions, which prevents their direct integration with state-of-the-art algorithms such as proximal policy optimization (PPO). In this paper, we extend the recently introduced robust safe gradient flow-based reinforcement learning (RSGF-RL) framework to overcome these limitations. Our formulation accommodates a general class of trajectory-level safety constraints, including chance constraints such as collision avoidance with a prescribed probability, and our analysis allows for biased value and gradient estimators. We show that, with high probability, the constraint violation throughout training is bounded by the estimator bias plus a term that vanishes as the number of rollouts grows, and decays to this level if the initial policy is unsafe. We also establish the convergence properties of the iterates to approximate KKT points of the original safe RL problem. Finally, we show that RSGF-RL implemented with PPO-style estimates enables zero-shot sim-to-real deployment of whole-body control policies on a Unitree G1 humanoid and collision-free waypoint navigation on a Crazyflie drone.
https://arxiv.org/abs/2611.19616
A frozen diffusion generator may need to change several output properties while preserving others. We study when guidance faithfully realizes such a structured request and whether local inverse correction improves its direction. Our account links this benefit to the interaction between the request and the objective-specific remaining-sampler Jacobian: conditioning alone is insufficient. Spectral steering combines a request-specific cosine-gain predictor with damped inverse correction and fixed-time relinearization for finite edits. Scalar nulls and matched-movement controls test the predicted mechanism. Across 8,923 instances in a retrospectively identified local regime, predicted and realized gains correlate at Spearman 0.79, versus -0.11 for condition number. Experiments span images, audio, robotics, and control and planning, including complex character motion generation. On 50 paired inpainting cases, four corrections improve unobserved-pixel PSNR over Adam by 3.25 dB at equal total derivative traversals. Together, the analysis and controls connect request-dependent geometry to directional gains and establish where these gains translate into downstream performance.
https://arxiv.org/abs/2611.19572
World-action models (WAMs) augment vision-language-action policies (VLAs) with future prediction so that the policy can anticipate how its actions will affect the physical world. Existing WAMs supervise complete future observations. However, observations typically change little over an action chunk, so the future loss is dominated by copying static background content rather than learning task-relevant robot and object motion, limiting policy learning. Moreover, existing WAMs rely on large video generation models or separate world models, incurring heavy training and inference compute that hinders closed-loop deployment. To improve policy performance, we propose visual residual learning, supervising only the feature change from the current frame to the future frame in a frozen DINOv3 space. Compared with complete future features, this residual learning target yields a higher feature rank with richer and more diverse representations of scene changes, and spatially concentrates on the robot arm and manipulated objects rather than the background. To further improve computational efficiency, we present ResWAM, a residual world-action model with 2.3B parameters in total, including a lightweight 0.5B dual-branch DiT that jointly trains an action branch and a residual branch via conditional flow matching. Without robot pretraining (i.e., pretraining on external robot action trajectories beyond the benchmark demonstrations), ResWAM achieves top average success rates on LIBERO-Plus (81.1%), RoboTwin 2.0 (93.8%), and RoboDojo (9.40% success, 13.41 score) at the lowest training (2.7-4.0 TFLOPs) and inference compute (1.0-1.4 TFLOPs), and leads real-robot Open, Put, and Stack (80.0% average, +10.0% over $\pi_{0.5}$) at 92.7 ms latency.
https://arxiv.org/abs/2611.19559
This paper proposes a novel approach for localizing the source of a three-dimensional scalar field using a team of heterogeneous cooperative mobile robots. Each robot measures the field intensity at its location and shares the information within the team. Measurement noise is explicitly considered, and heterogeneity arises from robot-dependent sensing accuracies, leading to heteroskedastic data. The proposed framework combines a weighted centroidal Voronoi tessellation-based formation strategy, defined on a spherical surface and parameterized by sensor noise variances, with a Weighted Least Squares estimator to compute the field gradient at the formation center. By explicitly accounting for sensor heterogeneity in both formation design and estimation, the method achieves robust gradient estimation under noisy measurements. In particular, we show analytically that this construction is nearly optimal. The covariance of the gradient estimate is derived in closed form, providing a lower bound on the attainable estimation variance for any formation on the sphere, and the proposed tessellation is shown to asymptotically satisfy the conditions under which this bound is attained. The estimated gradient is then exploited within a gradient-ascent strategy to guide the team toward the source location. Simulation results demonstrate the effectiveness of the proposed approach compared to alternative strategies and highlight its intrinsic ability to reconfigure online in response to robot failures, thereby enhancing its adaptability.
https://arxiv.org/abs/2611.19535
This paper presents a depth-aware path planning and collision avoidance framework for the safe navigation of robotic nursing assistants in cluttered environments with low-height obstacles. Conventional LiDAR-based navigation systems, while accurate for mapping and localization, fail to detect transparent or low-height obstacles, whereas vision-based systems are prone to noise and occlusion. The proposed framework integrates RGB-D depth sensing, rule-based obstacle avoidance, and the NavfnROS global path planner to drive the robot from an initial point to a goal point. The proposed approach is implemented on the Husky A200 robot, equipped with a Kinova Gen3 arm, to form a complete setup for a robotic nursing assistant. This system features a multi-layer architecture for global planning, perception fusion, and real-time control in the ROS-Gazebo environment. Experimental evaluations in a simulated cluttered environment demonstrate up to 40\% higher success rate and 35\% improvement in minimum clearance compared with single-sensor baselines. Furthermore, real-world experimental results validate the proposed approach as a computationally efficient and safety-aware navigation solution for autonomous healthcare robots.
https://arxiv.org/abs/2611.19534
Sequential composition of safety-certified controllers for hybrid and non-smooth robotic systems is well studied piecewise. Yet no existing framework reduces this question to a single algebraic condition. We construct a \emph{trajectory monad} for Filippov-regular differential inclusions. Its Kleisli category treats controllers as morphisms and sequential execution as trajectory concatenation. Control Barrier Function (CBF) certification, a guarantee of forward invariance, corresponds to a sub-monad of this category. Joint safety of composed controllers then follows from one algebraic condition. All controllers must land in a common sub-monad, which is closed under the monad multiplication. The sub-monad condition recasts recasts existing CBF composition practice categorically. It explains why a smooth intersection of multiple barriers supports safe composition, while enforcing barriers phase by phase does not. For the phase-by-phase case, we derive a closed-form Bernoulli model of the failure rate. We validate it over a randomized sweep of 3{,}000 instances. The same sweep shows that the standard baseline degrades under constraint conflict as the barrier count grows. This baseline enforces all barriers simultaneously as constraints in a single quadratic program (QP). We instantiate the framework on a four-mode piecewise-affine system and on a unicycle reduced-order model over heterogeneous terrain.
https://arxiv.org/abs/2611.19469
World models learned from offline data allow an agent to plan without further interaction with the environment, but the sampling-based planners used with them handle safety constraints only by scoring or rejecting sampled action sequences. We instead pose safe planning in the latent space of a learned world model as a successive convex program (SCP), in which a learned discrete-time control barrier function is imposed as a constraint at every step of the horizon. SCP requires smooth dynamics, so we build the networks from smoothed MaxMin activations and spectrally normalised layers, and bound the resulting linearisation error in terms of the Lipschitz bound and the smoothing width. As SCP is a local method, we start it from a sampled plan. On two Safety-Gymnasium navigation tasks with offline data, this planner has fewer violations than MPPI and CEM drawing 128 times as many samples at equal wall-clock time in our main comparison, though not significantly, and earns greater return at longer horizons, and is much safer than SCP without a sampled start. SCP converged more often on Lipschitz-bounded models. Unlike sampling, it also reports at each step whether its plan is a feasible, first-order optimal KKT point of the planning problem in the model, or that no feasible plan was found nearby. Our implementation is publicly available at this https URL.
https://arxiv.org/abs/2611.19466
A surveillance drone on a bandwidth-starved link cannot send the video its operator would need, yet the operator's questions are often not about pixels: did three people converge, when, and where? We study what should cross a link of 1 Mbps down to 1 kbps instead, and in what order. BPS, a behavior-primitive stream, sends semantics first and pixels later: it runs a detector and a tracker onboard and sends short records (25 bytes on average) of multi-frame events selected by a mission card of 16-125 bytes, ahead of positions and pictures. We measure it by what a ground-side language model answers from it, under a preregistered protocol with a held-out set, against H.265 read by the vision fine-tune of the same model, H.265 on which the ground runs our own perception, a stream of quantized tracks, and onboard captions. On the development set (84% of questions about convergence, mostly one location) the event stream answers 74.3% of questions at 0.13 kbps. A reader that receives nothing answers 57.4%, and the video reader 59.6% at 100 kbps and 64.7% on the source frames. When the ground instead runs our perception on the video, video matches the stream at 100 kbps, about 800 times the stream's rate, and falls behind at 64 kbps and below. A track stream matches it at 96 times the bytes and loses 0.03 at 17 times. On held-out data the result is weaker: with events taken from annotated tracks, an oracle, the stream beats H.265 at 100 kbps (+0.13), but with onboard perception it does not (+0.01 [-0.04, +0.06]) and is no better than receiving nothing. Sent alone through emulated links the event records answer 0.739 at 1 Mbps and still 0.724 at 0.3 kbps; an every-frame track stream holds down to 1 kbps and video with ground-side perception down to about 100 kbps. On a Jetson Orin Nano the unchanged pipeline runs at 4.5 fps and misses most events.
https://arxiv.org/abs/2611.19463
Vision-Language-Action (VLA) models have shown great promise for robotic manipulation by mapping multi-modal semantics to physical actions. However, this mapping inherently struggles to align these coarse-grained semantics with fine-grained temporal execution. It leaves VLA models with limited generalization and insufficient robustness in cluttered environments. To overcome this issue, we propose ActionUNet, an efficient multi-scale fine-tuning framework that enhances pre-trained VLA models with minimal computational cost. ActionUNet first constructs a lightweight temporal U-Net within the temporal-aligned action feature space to fuse hierarchical structural priors, effectively bridging the scale gap between semantics and temporal executions. Recognizing that multi-scale modeling can disrupt microscopic temporal continuity and cause mechanical oscillations, ActionUNet then employs a conditional SIREN as a continuous action decoder. Equipped with explicit second-order smoothness constraints, this decoder guarantees temporal continuity and reduces high-frequency motion jitter. By smoothing temporal discontinuities from multi-scale fusion, this continuous formulation reduces mechanical execution failures while preserving the base VLA model's generalization and manipulation robustness. Extensive experiments on RoboTwin 2.0 and LIBERO-Plus benchmarks, together with real-world hard evaluations, demonstrate that ActionUNet significantly improves {\pi}0.5 success rates by absolute 9.8%, 6.1%, and 11.4%, respectively, while also generalizing to the regression-based OpenVLA-OFT backbone, highlighting its effectiveness and efficiency as a fine-tuning strategy. Code and implementation details are available at this https URL.
https://arxiv.org/abs/2609.34982
In indoor 3D localization inside multistory buildings, an elevator or an escalator accelerates for a few seconds and then cruises at constant velocity for tens of seconds. Gait excitation disappears, the specific force degenerates into a constant gravity reaction, and instantaneous velocity becomes unobservable within a 1-2 s window. The output of short-window learned inertial odometry consequently degrades toward the conditional mean of the training set, and vertical displacement is systematically underestimated. We propose Local-Token Causal Attention (LTCA), a long-horizon architecture that compresses each 1 s inertial window into a 1 Hz motion token and sustains, over a bounded 60 s causal cache, the motion state established at the onset event. We further introduce a pluggable barometer branch that, when online, supplies a direct elevation constraint to the vertical channel through local differential height observations, and we train it with paired online/fallback conditional training so that the barometer also regularizes the representation shared by the two modes during training. On MSIO, our benchmark of nine buildings, the RoNIN and TLIO instances reach 3D trajectory errors (ATE-3D) of 2.83/4.18 m and 2.44/5.12 m on the Seen and Unseen splits, respectively, with the barometer online. With the barometer disabled at inference, the same weights fall back to pure inertial inference and reach 3.24/4.28 m and 2.64/5.23 m, still below every native short-window baseline. We further validate the framework on the public RoNIN and TLIO benchmarks.
https://arxiv.org/abs/2611.19387
Modern autonomous driving and driver assistance systems rely predominantly on cameras, radars and LiDARs for environmental perception. However, these modalities face different environmental limitations, including poor illumination for cameras and adverse weather for LiDAR. Acoustic sensing, particularly using microphone arrays, could provide complementary perception capabilities for autonomous driving, but has remained largely underexplored. In this work, we design a horizontally arranged microphone array that functions as the "ears" of a vehicle, enabling localization of nearby moving vehicles in a bird's-eye view (BEV) representation that captures both their relative direction and distance. Using this array, we collect a real-world driving dataset, AcousticScenes, which contains 2,117 static and 2,863 dynamic samples with synchronized spatial acoustic signals, LiDAR point clouds and 3D bounding box annotations. Building on this dataset, we propose AcousticBEV, a learning-based method for acoustic BEV object localization, which estimates the center locations of moving traffic participants in BEV from the microphone-array signals, conditioned on the ego vehicle's own speed. Experimental results demonstrate that AcousticBEV outperforms both classical signal-processing baselines and other learning-based methods, achieving F1 scores of $47.7 \pm 1.2$% on "Easy" scenarios and $45.1 \pm 1.4$% on "Difficult" cases. Our findings demonstrate that acoustic signals from a low-cost, vehicle-mounted microphone array can localize nearby moving vehicles in both azimuth and distance.
https://arxiv.org/abs/2611.19381
Vision-Language-Action (VLA) models have demonstrated strong capabilities in multi-task robotic manipulation. However, they still face limitations in control precision and execution stability during critical manipulation stages, such as object grasping, alignment and insertion. To combine general-purpose manipulation capabilities with task-specific fine-grained control, while enabling effective coordination between generalist and specialized policies throughout multi-stage manipulation tasks, this paper proposes MARS (Modular Atomic-skill Routing System), a dynamic routing framework for reinforcement learning-based atomic skills. MARS integrates a frozen generalist VLA policy with multiple independently trained atomic skill policies. Using a shared RL Token state representation and the reference action of VLA, specialized control policies are optimized through reinforcement learning for different cr`itical manipulation stages. For high-level policy selection, we develop a dynamic router that integrates temporal context and skill semantics. The router employs Mamba-based recurrent encoding and a historical feature window to capture state evolution during manipulation. Skill descriptions of candidate policies are encoded into semantic queries, and attention mechanisms are used to establish correspondences between skill semantics and temporal state features. Experimental results show that MARS achieves an average Macro-F1 score of 90.47% across four LIBERO task suites under fixed-trajectory evaluation, outperforming an MLP-based router by 7.08 percentage points. In the simulated Tube Transport task, the full-task success rate increases from 20.04% to 85.67%. These results demonstrate that stage-specific reinforcement learning and temporal state-aware dynamic policy routing can improve multi-policy coordination and overall task execution performance in multi-stage robotic manipulation.
https://arxiv.org/abs/2611.19324
We address tracking a moving object on a conveyor with a Franka Emika Panda robot arm, despite an imprecise digital twin, measurement errors, and noise. Our novel approach augments a loop-shaped inverse-Jacobian PI controller with a residual policy learned via Soft Actor-Critic in the digital twin. A delay-compensated Kalman filter fuses out-of-sequence camera measurements to provide latency-aware moving target position estimates. We compare the achieved performance to a performance bound for a class of model-based controllers derived from a frequency-domain sensitivity analysis, quantifying the minimum achievable error imposed by uncertain kinematic model mismatch. Simulation and hardware experiments indicate that our residual policy drives tracking error below the bound. Across conveyor speeds (27.9-41.8 mm/s), our method reduces mean tracking errors from 1.1-1.4 mm (baseline inverse-Jacobian PI) to 0.3-0.7 mm. These results show that learning residual corrections in a limited-fidelity simulator, combined with delay compensation, can substantially improve model-based control under modeling and measurement uncertainties while maintaining reliable sim-to-real transfer.
https://arxiv.org/abs/2611.19318
When a vision-and-language navigation policy generates an action sequence autoregressively, the robot waits for a chain of dependent predictions. A small discrete action vocabulary makes it possible to score all short-horizon plans in parallel. We introduce VLN-Flash, a training-free decoder that evaluates the complete action-prefix tree in a single forward pass. Branch-specific attention and matched token positions preserve the original policy's conditional action distribution for each prefix in exact arithmetic. We use the score margin between the stepwise-greedy prefix and its strongest alternative at each depth to decide how many actions to execute. The robot then executes the plan with the highest joint score at the chosen depth before observing again. On matched inputs, VLN-Flash achieves a nearly $3\times$ speedup in action decoding over native autoregression on Uni-NaVid and StreamVLN. Compared with the native autoregressive decoders, it improves success rate and success weighted by path length (SPL) on R2R-CE and RxR-CE, with success gains up to 2.7 percentage points. On a Unitree GO2, VLN-Flash improves navigation success and cuts the mean wait between actions by about 160 ms after subtracting network transmission time.
https://arxiv.org/abs/2611.19287
Long-horizon embodied tasks fail in ways that end-task success cannot see. HospitalBench is a diagnostic benchmark of ten hospital chores in NVIDIA Isaac Sim, each two unscored warm-up steps followed by a causal chain of five scored sub-tasks in one continuous simulation: the state one step leaves (an object pose, a choice, a fact seen once) determines what the next step must do, unlike benchmarks that place sub-tasks side by side. Vision-language agents act through skill-level tools from camera images. A scripted robot that makes only the calls a model gets is the solvability check of a chore. From each trajectory the benchmark reads where a chain breaks, what the agent did after a failed call, and how far its own position estimate drifts from the true one. This version reports first results only: 15 of 60 planned runs (three models at high effort, one run each on chores T1 and T10 from both start states and T3 from one), with scripted-robot proofs on three of the ten chores. Of the 15 runs, 2 passed all five sub-tasks. Runs still going after sub-tasks 1 to 5 were 15, 8, 4, 2, 2. Of 189 failed calls, 168 were followed by a different call and 111 by a call that did not fail. These are first results and not a full benchmark. We will release the scenes, task chains, scripted robot, evaluation harness and trajectories.
https://arxiv.org/abs/2611.19274
This paper shapes a swarm into a target pattern, such as a figure taken from an image, by controlling its spatial moments. The image is sampled into a reference configuration whose centroid and covariance set the position and principal spreads of the formation, while its higher-order moments, expressed in covariance-normalized (whitened) coordinates, encode the remaining shape. In contrast with geometric formation control, which prescribes distances, bearings, or relative positions between specific agents, this specification concerns the swarm as a whole and assigns no agent to a particular place. The key mechanism is orthogonalization in the swarm's own empirical inner product: projecting negative-gradient velocities onto the complement of low-degree polynomials sampled at the agents makes every higher-order correction preserve all lower-order moments exactly. The resulting hierarchical controller has lower triangular error dynamics and converges locally exponentially for any positive gains, so orders can be added without undoing those already achieved. A joint alternative regulates the higher moments together through a regularized Gauss--Newton step while preserving the centroid and covariance dynamics. We give bounds on the number of agents and sufficient conditions for regularity. Both controllers admit a distributed implementation that uses only relative positions and neighbor communication and retains local exponential convergence when estimation is sufficiently fast. In a distributed simulation on a sparse graph, 350 agents form an airplane from an unstructured swarm through moments up to order ten; the hierarchy reaches the full specification, while the joint law approaches it with smaller velocities.
https://arxiv.org/abs/2611.19272
Reasoning about what an action will cause requires internalizing how a system behaves. As conditions seldom stay fixed, this must often happen within a few interactions. Yet systems that look identical in a snapshot can respond differently to the same action, depending on physical conditions the snapshot hides. Context-based models capture this separation by summarizing recent transitions into a latent context, but do not specify how that context should change the equations of motion. Physics-informed neural networks build those equations into the model, but describe only a single system and cannot recognize when it has changed. We present Physics-Informed Causal Reasoning (PICR), which bridges the two. Only a context, re-inferred from recent interactions, is updated; it conditions a fixed physics-informed network that determines the robot's inertia and potential energy, so the robot adapts to new physical conditions from its own experience without retraining or a hand-specified robot model. On a Franka Panda arm in NVIDIA Isaac Sim, PICR predicts motion under changing and unseen payloads, reducing error on the latter by 1.6-3.8x over baselines, and extends to counterfactual queries under alternative torques. Our analysis shows that the context, inferred solely from interaction, encodes payload physical properties, tracks their changes, and governs the arm's response. The gravity and inertia terms induced by the context further demonstrate its physical validity, each agreeing with the simulator at the level of individual Lagrangian components.
https://arxiv.org/abs/2611.19251
Intent-driven exoskeleton control needs a continuous, causal estimate of the wearer's joint kinematics and an answer to a second question: whose data should train the estimator for a new user? Borrowing the user's recordings of other activities, or other users' recordings of the same activity, is usually evaluated as "more data" versus "less data," which confounds the information a source carries with the optimizer updates it adds and hides negative transfer behind averages. We separate these effects for hip, knee, and ankle angle estimation from high-density sEMG and an IMU across six locomotion activities on all ten participants of a public dataset. A compact residual causal TCN-GRU decoder is frozen, and an update-matched replay control prices every added source against the same number of updates on the user's own data. A user's own other activities raise $R^2$ from 0.722 to 0.788, of which +0.048 is information beyond replay; stair descent is a reliably harmful donor and sit-stand-walk a reliably beneficial one. Other users' same-activity data help while the user has only the target activity ($G_\mathrm{T} = +0.053$) and half as much once own other-activity data exist ($G_\mathrm{W} = +0.028$); at full calibration, donor value collapses after one to two target records. Per-target gains from the two sources correlate, and leave-one-participant-out gates recover most of the cross-subject value only at low calibration. Data value is hierarchical, state-dependent, and directional; the bottleneck is deciding, without test labels, when to transfer.
https://arxiv.org/abs/2611.19238
In simulation, Vision-Agent-Action (VAA) performs pick-and-place in robot cells with gripper or base-frame changes without robot demonstrations or fine-tuning. It uses pre-trained vision-language and language models for interpretation, measures the command map and camera geometry from its own motion and fingertips, estimates lens distortion, and plans grasps from measured object depth. A controlled study of eight public LIBERO-fine-tuned vision-language-action models (VLAs) varies robot hardware, camera geometry, appearance, and base frame. Gripper and frame changes dominate their losses in the tested fractional design. Across 18 cells with these changes, frozen VAA achieves 230/360 successes (64%), the refined VAA 268/360 (74%), and the strongest released VLA on this subset 76/360 (21%). Paired ablations support the self-calibration bundle, lens estimation under compounded changes, and the measured-depth pipeline. On a two-step LIBERO-Goal chain with object state carried between instructions, VAA completes 54/60 chains in changed cells, while four evaluated VLAs complete none.
https://arxiv.org/abs/2611.19198