Papers

Accurate neural world models are central to model-based robotics, where they enable robots to predict future states from previously observed trajectories. Multi-step autoregressive training improves long-horizon prediction, but fixed rollout horizons also increase computational cost and can amplify early training errors when the model is still inaccurate. Existing training schemes typically use the same rollout length throughout optimization, independent of the model's current predictive reliability. We propose an epistemic uncertainty-driven adaptive rollout strategy for offline world model training following an auto-curriculum training scheme. Instead of always unrolling to a fixed horizon, the model terminates autoregressive rollouts once epistemic uncertainty exceeds a threshold calibrated from a warm-up phase. We study two uncertainty estimators: a five-head ensemble with a shared recurrent backbone and Monte Carlo Dropout. A two-stage warm-up procedure stabilizes uncertainty estimates before we enable adaptive truncation. Experiments on ANYmal-D and ANT show that ensemble-based adaptive truncation matches or improves the prediction accuracy of fixed-horizon training and the RWM-U baseline while requiring substantially fewer cumulative rollout steps. Training a world model on ANYmal-D following the presented approach reaches comparable final performance with the baselines with roughly 72% less rollout computation. These results indicate that epistemic uncertainty is useful not only for downstream policy regularization, but also for making world model training itself more compute-efficient.

Fast-WAM shows that video-action co-training improves control without generating future video at inference, making the representation from a single video diffusion Transformer forward central to action generation. However, future-observation prediction does not explicitly prioritize the future dynamics and visual structure needed for control. We present MT-WAM, which retains the original training objectives and adds complementary supervision for future two-dimensional point trajectories and visual features. A lightweight dual-stream branch copied from the video backbone's final blocks provides target-specific processing, while a structured attention mask prevents cross-stream attention. Motion-stream tokens supply additional dynamics conditions to the action expert. Future visual-feature prediction provides supervision in a feature space that captures object and spatial structure. This supervision trains the video backbone to provide more informative visual context for action generation under changing visual conditions, without adding visual-feature-stream tokens to action conditioning. At inference, MT-WAM uses video and motion caches computed once per replan and skips future-video prediction. Without additional embodied policy pretraining, MT-WAM achieves 98.2% success on LIBERO and 73.7% on LIBERO-Plus, exceeding Fast-WAM by 23.8 percentage points on the latter. On RoboTwin 2.0 Clean2Rand, Random success increases from 6.30% to 19.40%; across four real-world tasks, average success increases from 67.0% to 77.8%.

Motion tracking can reproduce humanoid loco-manipulation from a single retargeted motion clip, but a policy trained on a fixed reference primarily reproduces its demonstrated transport outcome. Although the source trajectory visits intermediate object displacements, transport termination is demonstrated only at its endpoint. We identify this mismatch as the termination-versus-passage gap: intermediate displacements are observed as passage states rather than termination-complete outcomes. We introduce Distance-Conditioned Reference Recomposition (DCRR), which relocates the demonstrated termination segment to intermediate transport states. A frozen tracking teacher replays the recomposed references under closed-loop dynamics, and the retained trajectories are relabeled by their achieved object placements and distilled into a reference-free policy. This procedure constructs distance-conditioned supervision from the interaction behavior encoded in the source motion. Across Carry, Kick-Push, Crouch-Push, and Drag, DCRR-BC produces command-dependent transport with an overall normalized distance mean absolute error (MAE) of 0.15, compared with 0.28 for source-only behavior cloning. RL fine-tuning further improves the command response and execution robustness in the training simulator and under sim-to-sim transfer. Finally, hardware experiments demonstrate transport-distance modulation across all four interaction modes.

Embodied foundation models are constrained by the limited scale and diversity of robot demonstrations, motivating the use of large-scale egocentric human interaction data. However, how to effectively incorporate such data into embodied-model pre-training remains unclear because of substantial embodiment and action-space gaps between humans and robots. We present AtomEgo, a systematic study of ego–robot co-training supported by a curated corpus of approximately 2,659 hours and a scalable data processing pipeline. Across vision–language–action and world–action model architectures, we investigate three representative paradigms: joint co-training with domain-specific action heads, progressive ego-to-robot transfer through embodiment alignment, and joint video–action modeling. We evaluate these paradigms through multi-task real-robot experiments and language-conditioned cross-embodiment representation analysis. Our results reveal a simple principle: Data Scale * Alignment Quality –> Capability Gain; egocentric data can improve generalization, but their value depends on how effectively they are aligned and utilized. This principle can provide practical guidance for scalable ego–robot pre-training.

This work presents Robotic Multiphase Interaction (RMI), a setting in which liquid enters a porous material and interacts mechanically with its deforming solid skeleton. Manipulation can therefore change pore volume, expel or redistribute retained liquid, and alter grasp stability at the same time. Spilled liquid can also create safety risks in domestic and manufacturing settings. This differs from most manipulation of solid objects and from tasks that involve both liquid and solid while keeping the phases spatially separate. We study a sponge filled with water as the first RMI example. We use implicit incompressible porous flow with smoothed particle hydrodynamics as the dynamics engine and enable robotic manipulation by adding Coulomb contact memory, hybrid velocity and force regulation, and a stability gate for lifting. The resulting environment connects robot commands to changes in the coupled liquid and solid state. A world model conditioned on actions predicts how this state evolves under candidate commands, while a temporal UNet generates actions using either Diffusion Policy or rectified flow matching. Our world model reduces retained water prediction error by more than $60\%$ compared with the baseline. The best action sequence selected by the world model from policy proposals further reduces the predicted terminal water error by about half. These improvements show that modelling the coupled liquid and solid state helps the robot predict how its actions affect both the porous object and the liquid held inside.

Humanoid locomotion over complex terrain requires anticipating footholds that may no longer be visible at touchdown. Limited camera coverage and self-occlusion make it necessary to retrieve relevant terrain information from earlier observations. We present FootQuery, a perceptive locomotion framework that queries depth history using each foot's predicted next touchdown. The policy predicts touchdown locations and uncertainty from proprioception and uses these distributions, together with per-foot features, to query sparsely sampled historical depth frames. During training, realized contacts are projected into historical images to supervise retrieval at the regions where those contacts were visible. The retrieved per-foot features are fused with global visual memory to generate control actions. A progressive force-assistance curriculum supports early exploration, while event-consistent tread-midline shaping encourages coordinated stair contacts. Deployment requires only proprioception and onboard depth images. In simulation, the complete framework outperforms its component ablations on the most challenging tested stairs, gaps, and platforms. Real-world experiments on a Unitree G1 demonstrate continuous traversal with a single policy across outdoor stairs and indoor routes combining stair ascent and descent, platforms, and gaps. These results support organizing visual history around anticipated contacts for perceptive humanoid locomotion.

Collaborative robots are increasingly deployed in industrial scenarios characterized by frequent product changeovers. As an intuitive programming method, kinesthetic teaching facilitates rapid robot deployment. However, users may overlook the configuration of the robot during kinesthetic teaching, leading to degradation in operational performance. Operational performance refers to the capability of the robot to generate motion and can be quantified by the Minimum Singular Value of the Jacobian matrix. To address this issue, this paper proposes an online dynamic force guidance method that integrates performance constraint and optimization mechanisms. Specifically, variable admittance control maintains the operational performance of the robot above a predefined threshold, while a virtual force actively guides the user to drag the robot towards configurations with improved performance. Experiments are conducted on a 6-DOF collaborative robot, comparing three typical paths in the task space. To evaluate the quality of the taught trajectories, trajectory playback experiments are conducted to analyze the relationship between the operational performance of the robot and the work efficiency. The results demonstrate that the proposed method effectively enhances the operational performance of the robot and consequently improves the work efficiency, holding significant value for reducing production takt time in industrial deployment.

End-to-End autonomous driving models commonly predict future waypoints from sensor inputs and convert them into vehicle control commands through a downstream controller. However, conventional waypoint-based imitation learning mainly minimizes coordinate-level errors, making it difficult to capture scene-dependent path-speed changes and temporal instability across waypoint outputs. In this paper, we propose an offline teacher-signal generation and learning method for trajectory-output stabilization based on a Scene-Conditioned Interacting Multiple Model (SC-IMM) to mitigate this issue. The proposed method converts expert trajectories into path-speed states and performs IMM updates conditioned on scene cues to generate path-speed teacher labels and mode posterior probabilities. The generated signals are added to the original trajectory loss as auxiliary supervision during training, while the inference structure and waypoint controller remain unchanged. In closed-loop evaluation on 100 short routes in CARLA Town12, the proposed method improved the driving score by 28.0% and reduced Collision/km by 62.3% compared with the baseline, while also improving jerk and trajectory-variation metrics. These results demonstrate that offline teacher signals embedding scene-conditioned motion-model cues can guide trajectory-output driving models toward more stable closed-loop behavior.

Open-vocabulary 3D scene mapping aims to build a persistent representation of the objects in an environment. Existing systems typically rely on engineered mapping pipelines to associate observations, merge information across views, and maintain a consistent scene representation over time. Many additionally store feature-rich object representations, such as embeddings or image crops, increasing the size and complexity of the persistent memory. We introduce SceneLM, a Scene-Language Model that directly maintains a textual scene map. The full scene is represented as a structured text list of objects, which serves as the model's only persistent memory. For each input image, the model reads the current scene state and updates the map by adding, editing, and removing objects. To learn this behavior, we introduce supervision tasks for iterative scene map maintenance together with an automatic annotation pipeline that generates training data from images without human labels. We evaluate SceneLM on both a language-grounded retrieval benchmark and a localization benchmark. Across both benchmarks, the model produces a scene map that achieves competitive performance with complete mapping systems built from dedicated perception and geometric modules while producing a scene representation that is 6-12x more compact. We further show that SceneLM can be run online on an edge device through experiments on a quadruped. These results show that a persistent open-vocabulary 3D scene map can be maintained directly by a single vision-language model using only a lightweight text representation. Training and inference code is available on https://goldengait.github.io/scenelm/.

Underwater robot development is often hindered by the complexities of waterproofing and wiring, which significantly delay the rapid prototyping process. This paper presents MarineCraft, a modular toolkit designed to accelerate the development cycle through structural reconfiguration. The system features self-contained, waterproof propulsion modules that integrate power, wireless communication, and actuation. By eliminating centralized wiring and the need for repeated sealing, MarineCraft allows diverse robot geometries to be assembled and tested in minutes rather than days. Experimental results demonstrate that this reconfigurable architecture enables fast, iterative design cycles while maintaining reliable operation and leak-free performance at depths of up to 2.5 meters. Our toolkit effectively lowers the barrier to underwater robotics by transforming modularity into a vehicle for rapid physical prototyping.

Robotic fabric comparison needs to actively combine visual appearance and tactile cues. Here, we present AVT-Fabric, an RGB-first framework that allocates tactile evidence according to the difficulty of each comparison. A dual-scale gate evaluates answer-token confidence and raw logit separation to determine whether another force-tagged GelSight observation is needed. Compact textual memory preserves the executed history, and majority voting consolidates the selected predictions. On 400 held-out comparisons, AVT-Fabric achieves 98.0% accuracy with a compact 7B Multimodal Large Language Model (MLLM), surpassing the 94.0% reported by the 90B MLLM-Fabric baseline by 4.0 percentage points while processing only 1.60 of five available stages on average. It improves on matched passive inference by 9.25 percentage points and reduces model-side latency by 61.8%, while also improving on RGB-only accuracy. Four additional MLLM backbones support the generalizability, accuracy, and efficiency of the framework. This framework is also deployed on a real robotic system, achieving 78.1% pairwise ranking accuracy and correct fabric selection in seven of eight application scenarios. AVT-Fabric demonstrates that adaptive evidence selection can improve both the accuracy and efficiency of robotic visuo-tactile reasoning.

This paper presents a comprehensive framework for robot manipulation failure analysis that includes binary failure detection, failure categorization, explanation generation, and the additional capability of failure onset localization, which aims to identify the earliest moment at which a robot execution deviates from a valid task-completion trajectory and is ultimately followed by task failure. To address these tasks, we propose ProTracer, a training-free framework that leverages existing Vision-Language Models (VLMs) together with proprioceptive signals for failure analysis. Our method uses proprioceptive dynamics to identify temporally informative action boundaries and converts richer robot-state signals into structured natural-language descriptions that can be jointly analyzed together with visual observations by the VLM. This design combines the temporal precision of proprioceptive signals with the multimodal reasoning capabilities of modern VLMs without requiring additional model training. We further introduce FailTime, a benchmark with synchronized visual and proprioceptive observations for evaluating conventional failure diagnosis tasks as well as failure onset localization. Experiments demonstrate that ProTracer achieves strong performance across both conventional failure diagnosis tasks and the newly introduced failure onset localization task, highlighting the importance of proprioceptive reasoning for fine-grained temporal failure analysis.

Automated needle-loop hooking is a critical step in flexible microelectrode (FME) implantation. This paper presents MicroHookACT, an imitation learning-based visuomotor policy for automated 3D hooking under monocular microscopic vision. First, a unidirectional hooking strategy exploits defocus cues and optical-axis guidance to enable palpation-free precise alignment and contact-rich threading. Second, an action-supervised object attention module built on a frozen ViT backbone learns to focus on the micro-needle tip and micro-loop directly from human demonstrations, without requiring manual visual annotations for training. Third, attention-centered global coarse and local fine features are dynamically weighted according to predicted action progress, enabling a single ACT policy to adapt to changing defocus blur and visual requirements throughout the operation. In the experiments, visuomotor policies were trained on 60 human demonstrations and evaluated under five setups with varying difficulties. Our MicroHookACT framework achieved the highest overall success rate of 96.7% with an average execution time of 11.5 s. These results demonstrate the potential of visuomotor policy learning for micron-level control under varying operating conditions.

Vision-Language-Action (VLA) models pre-trained on large-scale, closed datasets have demonstrated remarkable success across diverse robotic manipulation tasks. However, their long-term real-world deployment necessitates continuously acquiring new skills while retaining previously learned capabilities. While pioneering works have explored continual VLA adaptation using techniques such as experience replay and reinforcement fine-tuning, they overlook a foundational mechanism: action normalization, which determines the underlying coordinate system in which policies perceive and execute physical actions. To bridge this gap, we systematically evaluate five normalization strategies across four real-world task streams covering single-arm and bimanual manipulation. Our analysis reveals that existing protocols induce severe failure modes due to inter-task coordinate drift, limited motion coverage, or train-test coordinate mismatches. Motivated by these insights, we formulate three core design principles: consistency, coverage, and causality (3C), and introduce foresight action normalization (FAN). FAN estimates normalization statistics once from a small, task-independent calibration set prior to continual learning and freezes them throughout adaptation. Across all evaluated streams, FAN achieves the highest performance and demonstrates consistent robustness, providing insightful guidance for building stable action representations in achieving effective lifelong VLA adaptation.

Event cameras detect per-pixel brightness changes asynchronously on microsecond timescales, with high dynamic range and low power draw. These are desirable properties for robots that move fast or work in difficult lighting conditions. However, integrating an event camera into a real-world robotic pipeline still requires substantial effort: plug-and-play drivers do not exist, event streams are recorded in a variety of incompatible file formats, and most projects rely on custom research-grade code. Here, we present EventCV, an open-source and extensible Rust library with OpenCV-style Python bindings that lowers the entry barrier to working with event cameras. EventCV provides a wide range of features: denoising filters and geometric transforms, augmentations, corner detection and unsupervised feature learning, contrast-maximization motion estimation, a video-to-events simulator, and Open Neural Network Exchange (ONNX) inference for deployment in robotic stacks. EventCV integrates the Neuromorphic Drivers package, allowing an event camera stream to be processed directly in real time. No existing toolkit covers this range of operations in one package, and EventCV builds representations and decodes files 1.1x to 3.7x faster than the currently available libraries. We deploy EventCV on a Jetson Orin AGX and present three robotics case studies spanning object detection, on-device model inference, and localization. Project webpage: https://eventcv.net.

Physical control tasks in the natural world, such as driving or object manipulation, frequently exhibit dramatic variations in sensory and compute complexity over time. Correspondingly, a natural resource-efficient choice for robot control is to dynamically switch between control modes with varying resource allocations. However, such "mode-switching controllers" (MSCs) have historically required laborious, expert-driven design and synthesis for each new task. Driven by these design difficulties, modern robotic control architectures often fall back to a wasteful "monolithic" one-size-fits-all structure, where resource allocation is permanently anchored to the hardest, most resource-intensive task phases. To facilitate the design of performant yet efficient MSCs, we propose LLM-Guided synthesis of Efficient Mode-Switching Control Architectures (LEMCA). LEMCA represents MSC designs as interpretable programs to be iteratively refined in an evolutionary loop. To evaluate design fitness, we propose MSC-compatible extensions of automated controller synthesis approaches, such as reinforcement learning in simulation. LEMCA then leverages the semantic priors, reasoning, and coding capabilities of Large Language Models (LLMs) to iteratively edit controller modes, their corresponding sensory-compute resource allocations, and mode transitions. Our experiments across diverse control benchmarks show that LEMCA consistently discovers strategies that surpass the Pareto frontier of monolithic designs by reclaiming wasted resources during "easy" task phases. LEMCA thus presents an automated, low-effort path to synthesize resource-efficient MSC designs.

Human-readable maps provide an intuitive interface for specifying robot destinations, but connecting their schematic geometry to egocentric observations remains challenging. We present NaViRRator, a framework that translates user-specified start and goal locations on such maps into navigation instructions for a pretrained vision-and-language navigation (VLN) policy. Its core method, RouteScribe, separates route inference from verbalization by first generating an explicit route scaffold in map-image coordinates, which a pretrained vision-language model (VLM) converts into a navigation instruction. We construct the scaffold with start–goal line conditional flow matching (SGL-CFM), which deforms a straight start–goal waypoint sequence into a map-conditioned route. During execution, the VLN policy receives only the instruction and egocentric observations, while the map and scaffold remain upstream, allowing executor replacement without retraining the map-to-language modules. Real-world experiments show higher success rates and success weighted by path length (SPL) than direct map-to-instruction generation, A*-based scaffolding, and Gaussian-source conditional flow matching. Qualitative results further show clearer salient turns and better preservation of the intended maneuver sequence, supporting route-grounded language as a modular interface between human-readable maps and pretrained navigation policies.

Although conventional controllers and disturbance observers (DOBs) are the standard for precision tracking in manipulators, they suffer from parameter uncertainty, nonlinear friction, and compound disturbances. This study proposes a residual reinforcement learning DOB framework that pairs an analytical observer with an RL policy. The deterministic baseline operates within a reliable region, whereas the RL policy explicitly targets the residuals that the model cannot capture. To make this compensation disturbance-aware, an estimator network aligns the observation history with a privileged disturbance context, organizing the latent space by disturbance regime and enabling rapid adaptation across disturbance transitions. To guarantee stability, we derived and enforced a state-dependent action bound on the RL policy from an input-to-state stability (ISS) analysis such that the closed loop provably confines the tracking error to a certified envelope for arbitrary policy outputs. Experiments on a 6-DOF manipulator demonstrated consistent improvements in disturbance estimation and tracking, including a 27.8% tracking-error reduction on real hardware under zero-shot sim-to-real transfer and a 38.0% reduction under a base-vibration disturbance that was not observed during training.

We introduce Learned Optimal Inverse Kinematics (LOInK), a method to generate approximately optimal solutions to an inverse kinematics problem. When trained on data consisting of sampled configurations and associated task variables and a given cost function, LOInK learns a bi-Lipschitz invertible mapping from configuration space to a decoupled task/latent space, and moreover, the latent space is structured so as to place cost-minimizing solutions at the origin. This enables efficient sampling of cost-minimizing solutions via a network-inversion algorithm based on operator splitting. We demonstrate the proposed approach on three problems: an illustrative three degree-of-freedom manipulator problem; a quadrupedal climbing robot for which LOInK can generate near-optimal solutions on average 31 times faster and up to 100 times faster than a constrained optimization approach; and a simulated soft actuator as a purely data-driven example, in which LOInK can explicitly generate high-quality solutions, unlike existing generative approaches that require diverse sampling and evaluation of candidate solutions.

Vision-language-action (VLA) models map visual observations and natural-language instructions to robotic actions, but distribution shifts can compromise their reliability. Because these models may still succeed under out-of-distribution (OOD) conditions, detecting OOD inputs alone is insufficient to predict execution failure. In this paper, we introduce VLA-Scope, a two-stage framework that combines input-shift characterization with execution history to predict failure during OOD rollouts. The first stage uses pooled image and language representations to detect OOD inputs and classify their shift categories. For inputs flagged as OOD, the second stage combines the predicted category, action-prefix features, and execution progress features. A logistic regression model shared across shift categories updates failure risk as execution proceeds. We evaluate the framework with OpenVLA on ten LIBERO-Spatial tasks using leave-one-group-out cross-validation. OOD detection achieves a ROC-AUC of 0.9454, and shift classification achieves 91% accuracy. Evaluated independently of the OOD gate on all 1,400 OOD rollouts, the failure predictor achieves a ROC-AUC of 0.8497 after 60 executed actions, compared with 0.7906 without execution progress features. It also achieves a higher ROC-AUC than the evaluated ActProbe and SAFE-MLP baselines. These results suggest that combining action features with temporally aggregated execution step representations improves failure prediction under input shifts.

Learning robot manipulation policies typically requires substantial demonstration data, which are costly to collect on real robots. Recent methods generate robot demonstrations from human videos by adapting recovered motion and validating the resulting trajectories in simulation. However, methods centered on motion-reference adaptation can limit behavioral diversity by retaining the demonstrated contact strategies and subtask orders, while insufficient understanding of task requirements and scene relations can reduce demonstration generation efficiency by generating invalid candidates. To address these limitations, we propose KnowDemo, a framework that uses structured manipulation knowledge from human videos to generate diverse robot demonstrations for a target workspace. To distinguish task requirements from demonstration-specific choices, we develop a knowledge extraction and reasoning module based on a vision-language model (VLM) that associates object and action descriptions with inferred task conditions, demonstration references, and permissible execution variations. To translate this knowledge into executable demonstrations, we resolve the descriptions against target-scene entities and geometry to guide candidate generation and screening before motion planning and simulation. The resulting demonstrations exhibit multimodal behavior through alternative contact strategies and valid subtask orders, with structured execution labels. Experiments demonstrate additional verified execution modes beyond a reference-only configuration and improved candidate planning success through task-guided grasp sampling. To validate the generated data for policy learning, we fine-tune the pretrained $\pi_{0.5}$ model on simulation data, achieving sim-to-real transfer across three tasks. Project page: https://zhiyuan-gao.github.io/knowdemo/

Vision-language-action (VLA) models built on pretrained vision-language models have demonstrated strong performance across diverse robotic manipulation tasks. However, VLA models that directly map current 2D observations to actions often lack sufficient spatial and temporal understanding, limiting their performance in precise and long-horizon manipulation. Recent methods enhance VLA models through geometric supervision and future-state prediction across the entire scene. However, these methods can suffer from redundant scene information, distracting the model from learning the geometry and dynamics relevant to the current interaction. To address this issue, we propose FOCAL-VLA, a framework that combines subtask-guided geometry distillation with implicit world modeling to learn representations of current spatial structure and future interaction dynamics. To focus geometric learning on the current subtask, we transfer geometric knowledge from VGGT to the VLA model by aligning geometry latents with features from subtask-relevant image regions. To capture the future 3D evolution of the current interaction, we incorporate implicit world modeling using Track4World features from current and future demonstration frames. The two complementary representations jointly guide action generation without running VGGT or Track4World at inference time. Experiments show that FOCAL-VLA outperforms baselines on both simulation benchmarks and real-world manipulation tasks. Project website: https://zhiyuan-gao.github.io/FOCAL-VLA/.

Quadrotors are increasingly deployed in applications such as agriculture, infrastructure inspection, and maintenance. In each of these applications, the robot must navigate complex scene geometry while remaining strictly collision-free. Unlike in ground domains, even minor collisions for aerial vehicles can result in the loss of the robot. This safety requirement induces a pair of technical challenges. First, the environment must be represented with sufficient fidelity to encode complex structure, even when no ground-truth obstacle data is available. Second, a motion planner must leverage this representation to determine a collision-free path to the goal. This paper proposes a system that addresses these complementary challenges. The proposed method, AirSplan, adopts a normalized variant of 3D Gaussian Splatting that encodes high-fidelity scene geometry. It then applies a novel reachability-based motion planner that leverages the differential flatness of quadrotors to compute continuous-time collision constraints that tightly overapproximate the robot's occupancy. Experiments demonstrate that AirSplan successfully finds a path in 81.2% of challenging test cases, a significant improvement over the nearest baseline method's 51.2%.

Vision-language-conditioned robot policies integrate perception, language understanding, and control for general-purpose manipulation. However, existing evaluations often focus on task success, isolated physical constraints, semantic refusal, or realized physical damage, providing limited insight into where safety fails during closed-loop manipulation. We introduce SafeStage, a lifecycle-structured benchmark for evaluating manipulation safety before, during, and after task execution. SafeStage contains 97 purpose-built risk scenarios organized into three stages. Initial-State Hazards captures safety-relevant relations that must be resolved before manipulating the target. Execution-Time Safety evaluates unsafe contacts, trajectories, region entries, and object interactions during execution. Final-State Hazards capture unstable or otherwise unsafe conditions remaining after nominal task completion. The benchmark evaluates realized interactions using event-based and state-based checks and reports native task success independently from stage-specific safety outcomes. We evaluate representative direct-action Vision-Language-Action (VLA) policies and policies with world-model-based policies under a common closed-loop protocol. Our results demonstrate that nominal task completion frequently coexists with safety violations and that different policies exhibit distinct failure profiles across the three stages. By separating task success from safety and localizing when violations occur, SafeStage provides a unified diagnostic testbed for evaluating and improving vision-language-conditioned robot manipulation policies.

Perceptual planning tasks require two key capabilities: accurately perceiving uncertain scenes and planning valid action sequences following logical rules. Conventional methods convert perception into discrete symbolic facts and then plan, discarding perceptual uncertainty and severing task-level feedback to perception. We introduce a generic, fully differentiable neuro-soft-symbolic framework that connects visual perception and task planning within a single computational graph. The framework maintains a continuous soft symbolic state, lifts domain rules into a differentiable soft-$T_P$ transition operator, and optimizes action logits over a short planning horizon. Gradients from the planning objective can also update the perception parameters, allowing task-relevant perceptual representations to be refined during planning. On Blocksworld, our method solves 40/40 LatPlan-40 tasks and 596/600 PlanBench-600 tasks, compared with 33/40 for LatPlan and 587/600 for the reasoning-model baseline, while requiring substantially less computation and time. In the perceptual-uncertainty ablation, our method improves the success rate from 59% with frozen perception to 83%. We further conduct task-and-motion simulations on Blocksworld scenes, providing an execution-level validation of the compatibility between decoded task plans and downstream robotic motion execution.

Generative robot policies can represent diverse, multimodal behaviors, but adapting pretrained policies to deployment-time constraints such as collision avoidance and orientation maintenance remains challenging. Existing inference-time steering methods typically apply gradient guidance through iterative diffusion or flow processes, which can be computationally expensive for real-time control. We propose INSPO, which formulates inference-time steering of one-step generative policies as trajectory optimization in the policy's input noise space. By optimizing the input noise while evaluating constraints on the induced state trajectory, INSPO searches the policy-induced behavior space without directly modifying generated actions. The optimization includes a regularization term that encourages solutions to remain consistent with the policy's input distribution and is solved online using population-based particle optimization. We evaluate INSPO on state- and image-based task-specific policies and generalist vision-language-action policies across Push-T, Can pick-and-place, and LIBERO-Spatial. INSPO improves task success and constraint satisfaction over best-of-N sampling and action projection, while comparing favorably with gradient-guided generation at lower runtime.

Robot localization is an ongoing challenge that demands mapping and positioning systems that are tolerant to viewpoint change. Event cameras are attracting increasing interest and adoption in robotics; however, dealing with viewpoint variance is an under-investigated problem in existing event-based localizers. In addition, event-based datasets that emphasize viewpoint variance for challenging localization situations are scarce. Here, we introduce an event-based visual place recognition (VPR) system that performs robustly under viewpoint changes. We converted five large-scale geo-tagged datasets, conventionally used to train frame-based localization systems, into synthetic event streams using Image-to-Event (I2E) conversion, and used them to fine-tune a pre-trained event-based vision transformer backbone with a multi-loss function, yielding a system we call MegaEvent that learns viewpoint-robust features for place recognition. We achieved an average Recall@1 of 82% across three existing event-based localization datasets, leading the next best event-based method by 20 recall points, and frame-based VPR models applied directly to event frames by 8 to 26 recall points. We introduce a new, challenging dataset - Springfield-Event-VPR - which features a 3.7km walking route recorded in three camera orientations for a total of 11.1km, which MegaEvent outperforms the strongest baseline by 9 recall points. The code for MegaEvent is available at https://github.com/AdamDHines/megaevent.

Vision-language-action (VLA) policies based on flow matching generate action chunks through repeated evaluations of an action expert. Increasing the number of integration steps raises inference cost, but does not necessarily improve closed-loop success. We propose Coda, which reallocates part of this integration budget to a single learned endpoint correction. A frozen policy first completes a few-step noise-to-action trajectory; a lightweight Transformer then predicts a demonstration-supervised residual using the candidate action, source noise, and shared observation-prefix cache. Only the corrector is trained. On 50 RoboTwin Easy tasks, five-step Coda improves success from 71.64% to 74.68% over the matched five-step baseline, while reducing forward latency by 30.2% relative to the default ten-step policy. A two-step configuration achieves 71.88% success with a 2.12$\times$ speedup. An independent 13-task control shows a 5.69-percentage-point gain at nearly equal latency, supporting correction as an effective alternative to additional integration. The same design also improves frozen official SmolVLA, raising two-step success from 60.8% to 69.4%. These results show that endpoint correction improves the quality-latency trade-off of frozen flow-matching policies.

Learned navigation policies typically consume observations as a temporally ordered history, with positional encodings tying each observation to when it was seen, making it difficult to reuse experience from earlier traversals of an environment. Systems that do reuse such experience usually construct an explicit representation, such as a map or a topological graph, and plan on it. We propose VNT-PA (Visual Navigation Transformer with Pose Attention), a transformer planner whose context is a set of depth keyframes indexed by camera pose. With camera poses as positional encoding, attention depends on the pose differences between keyframes rather than on their temporal order. VNT-PA is trained to imitate a shortest-path planner operating on the ground-truth scene mesh, predicting actions by querying the spatial context with only its current pose and the goal position. On point-goal navigation in HM3D validation scenes, VNT-PA reaches 93.3% success and 90.4% success weighted by path length (SPL), outperforming baselines that encode the same context as a temporal sequence or treat pose as an input feature, in both navigation performance and training efficiency. Because the spatial context is a pose-indexed set, frames from different trajectories can be fused at test time. The planner also degrades more gracefully under localization noise than a conventional baseline which plans on explicit maps. These results show that pose-stamped experience can serve directly as the environment representation for a learned planner, and that making attention depend on pose differences, rather than on temporal order, speeds up training and improves long-horizon navigation.

Collision-free motion planning requires reliable collision models from sensed environments and validation of states along a continuous trajectory. To make this tractable, most planners check for collision at discrete states along continuous trajectories against a single determinized model of the environment, introducing a trade-off between safety and computational efficiency. While continuous collision checking approaches that approximate the swept volume of the robot exist, they are computationally expensive or overly conservative. Data-driven approaches can learn the swept volume; however, these neural models are susceptible to approximation errors and are therefore often limited to serving as coarse filters for downstream collision checkers. In this work, we propose to learn a signed distance function of the swept volume as a probabilistic field, enabling quantification of epistemic uncertainty, incorporation of perception noise, and eventual integration into a chance-constrained trajectory optimization framework. We demonstrate our approach on challenging high-dimensional manipulation problems with significant sensor noise, both in simulation and on real hardware.