Publication Radar.

This paper introduces OrbitalGym, a simulation environment for studying when planning algorithms or system design determine the outcome of adversarial orbital encounters. In an orbital game, two non-cooperative teams of satellites maneuver during rendezvous and proximity operations to win a differential-reward objective. Advances in terrestrial game playing have prompted many methods for orbital games, but studies typically evaluate one method in one scenario, often against a passive or random defender. It is unclear whether better planning matters once vehicle capability (maneuverability, sensing, communications) and adversarial responses are considered. OrbitalGym is a JAX-based environment supporting varying numbers of agents, state representations (position, velocity, attitude, fuel), explicit or implicit communication, and proximity or conical field-of-view sensors. It compiles games to batched GPU code, enabling large-scale parallel simulation of policies across many configurations. This identifies regimes where policy choice affects outcomes and regimes where system capability dominates. We study two close-range games: pursuit-evasion, where one spacecraft tries to capture another, and lady-bandit-guard, where a guard prevents an attacker from reaching a protected asset. We compare PD, LQR, model predictive path integral control, and Monte Carlo tree search over ratios of thrust and propellant, initial separation, game duration, sensor capability, and the definition of capture. In pursuit-evasion the propellant ratio decides most encounters: against five of six evader policies, pursuer policy changes capture probability by 10 percent or more in at most 15 of 90 capability configurations. In lady-bandit-guard, policy matters more: against bandits able to reach the asset, a better guard policy raises defense probability by at least 10 percent in 56 to 74 of 90 configurations.

Human-level dexterity remains a central open problem in robotics, yet few simulation frameworks cover the full workflow of dexterous manipulation, from setting up a robot to collecting dexterous demonstrations and evaluating policies. We present DexSuite, an all-in-one simulation framework for dexterous manipulation built on the Genesis simulator. DexSuite provides 30 tasks and assembles its 10 arms and 12 end-effectors, from parallel-jaw grippers to dexterous hands, into more than 100 single-arm robots and thousands of bimanual pairs. It also ships a standardized environment API, integrated teleoperation, and simulated tactile sensing. Moreover, we define a benchmark of 16 task-embodiment pairs spanning diverse contact complexities and horizons, and release a dataset of 4,700 successful episodes in simulation from human teleoperation, scripted experts, and learned policies. We release DexSuite as open source and hope others will build on it. Videos and datasets are available at https://dexsuiteorg.github.io

Motion retargeting maps a demonstrated human movement onto a robot to compute its joint motion. This study examines how the construction of the desired wrist position affects whether computed arm joint angles reach their permitted limits on a Unitree H1-2. A reconstructed target is formed from demonstrated arm directions and robot segment lengths, so its shoulder-to-wrist distance depends on the demonstrated angle between the segments even when no posture-matching preference is applied. A reach-normalized target retains that direction while rescaling the distance, using the same processed demonstrations and the same joint-angle method. Across a forward-reach sweep the reconstructed target places a joint angle at a bound in 10 of 20 repetitions and the reach-normalized target in none. Across 25 recordings from one participant the corresponding counts are 8 and 6 on the reaching arm and 14 and 5 on the supporting arm, and a second participant repeats the bricklaying and lifting comparison. In hardware replay a joint is commanded to a bound in 12 of 16 direction-matching arm-records and in none of the 16 normalized records, and within the eligible records both tracking error and executed torque at those frames exceed their values elsewhere in the same run. Cross-body reach and arm-to-arm clearance identify limits of radial adjustment. The findings support treating target construction and posture matching as separate retargeting decisions, and evaluating both alongside the wrist positions and clearance that manipulation requires.

Humans navigate their environment based on their vision system, anticipating how the ground will respond before committing to a step. Self-supervised learning could provide legged robots the same ability, but the supervisory signal is typically chosen on intuition, and its effect on what the model learns is rarely examined. We present Terrain-Aware COntact Supervision (TACOS), which proposes a meaningful and consistent footstep instability score, based on the residual of a factor graph in legged state estimation. It projects the continuous score into preceding images to supervise a dense visual predictor. We propose two variants, a hard score that exploits hindsight and a causal soft score available immediately. Both metrics are compared against state-of-the-art alternatives in two independent evaluations that share no data or labels. On 1716 manually annotated contacts, both variants outperform every evaluated alternative at detecting instability and ranking its severity, improving pooled severity ranking by 28 % over the strongest baseline. On three outdoor test runs, our metric also trains the most accurate visual predictor, improving rank correlation by 58 % over the strongest alternative, and by a further 25 % when supplemented with onboard context. In a real-world experiment over a 50 m forest trajectory, the trained predictor handles non-geometric terrain properties such as mud and vegetation, while stepping around geometric obstacles like fallen branches. Code is available for reproducibility at https://github.com/norlab-ulaval/TACOS.

Robots, autonomous vehicles, and consumer 360° cameras use fisheye lenses for wide fields of view (FoVs). Simulation-ready digital twins require photorealistic rendering across the full FoV and an aligned collision mesh. However, prior fisheye Gaussian Splatting methods only rasterize up to 180° and do not produce meshes, while Gaussian surface reconstruction methods assume pinhole training images. We present GOFish, the first 3D Gaussian Splatting surface reconstruction method to train directly on fisheye images, rasterize FoVs beyond 180°, and extract an aligned mesh from the same scene. Calibrated densification recovers the surface across the entire FoV, while Calibrated Mip Filtering and Wide-Angle Appearance Modeling keep pixel blur and lens falloff out of the scene and mesh. Across the 200° FIORD and 205° FullCircle benchmarks, GOFish improves PSNR over the best prior method by 2.28 and 1.97 dB, respectively. It also outperforms fisheye rendering baselines within their own FoVs and achieves the highest mesh F-score on FIORD and the main ScanNet++ set. Together, the Gaussian scene and mesh form a LiDAR-free, simulation-ready digital twin that renders in real time, ray traces the full FoV, and provides collision geometry.

Scaling visual reinforcement learning (RL) with GPU-accelerated simulation is computationally expensive due to rendering and visual-policy inference costs. Consequently, two-stage teacher--student distillation is commonly used to train visual policies, but introduces an imitation gap. In this preliminary study, we introduce \echomethod{}, \textbf{a single-stage off-policy RL framework that scales experience collection without proportionally increasing visual computation}. Through \textbf{inverse distillation}, a state-based auxiliary policy imitates the evolving visual policy and collects inexpensive, unrendered experience. Echo leverages this experience for off-policy value learning, directly improving the visual policy through a shared privileged critic. We instantiate \echomethod{} with FastTD3 and evaluate it on three visual control tasks and two dexterous manipulation tasks in Genesis. Relative to \sdpgmethod{} under the evaluated configurations, Echo reduces time to reference returns by an average of \textbf{60\% across the three locomotion tasks} and \textbf{56\% across both hand tasks}, while improving peak mean returns by \textbf{62\% and 12\%}, respectively. Both hand tasks also achieve higher mean returns at matched simulator-transition budgets. These initial results suggest that scaling off-policy experience is a promising approach to efficient direct visual RL.

On PushT, our fixed action updater (1,505 parameters) solves 8 of 16 scenes reserved from our configuration selection. The comparison planners perform as follows: Selected Adam (7 tasks solved, total planning and encoding time 218.440s), adapted iCEM (2 tasks solved, total planning and encoding time 49.309s), released LePlanner (16 tasks solved, total planning and encoding time 5.652s). The total planning and encoding time for our action updater is 59.079s. Note that our action updater does not satisfy the task-cost criterion we fixed before evaluation. We also conduct another matched experiment where we only replay the first ten commands in each context and then allow the next solve 4, 8 or 12 updates (subsequent solves all get 8 updates). Among all 16 contexts (8 scene clusters), the 3 budgets get 9, 8 and 7 successes, respectively. Interestingly, the extra updates did not rescue any failing task in this experiment. All the changes in success/failure happen in one scene cluster, where we observe that the extra updates decrease the costs of the selected latent world-model trajectories for our updater, while the tasks that used to be successes become failures. The code, weights, evaluators, execution traces and cost accounting accompany the paper.

Vision-Language-Action (VLA) models offer a promising paradigm for general-purpose household robotics. However, inference and data-collection strategies that have been deployed in smaller-scale settings do not necessarily generalize to large-scale, diverse task suites. To address this gap, we conduct a systematic, paired evaluation of previously proposed inference and data-collection strategies across all 100 tasks. Our evaluation uses a fine-tuned $π_{0.5}$ model based on the 2025 BEHAVIOR Challenge winner, which achieves a Q-score of 0.39 on the public test set. These results show that proposed strategies do not transfer uniformly to new environments and tasks. Project page, videos, and detailed analysis: https://rubenhammele.github.io/behavior2026

Radar systems can recover scene geometry and recognize predefined categories, but these outputs do not provide a persistent spatial representation for open-ended language queries. Such queries require linking observations from different viewpoints to the same physical locations. We present TextRadar, a framework that constructs a scene-level 3D occupancy-language field from panoramic radar observations. The field combines occupancy with scale-dependent language features at surface locations, allowing the reconstructed scene to be queried with new descriptions. A single-frame radar distillation model predicts aligned ranges and language features from each radar observation. Using radar-estimated poses, we register these predictions to reconstruct the field from radar alone in a new scene. Experiments on approximately 20k synchronized observations across seven buildings demonstrate scene-level natural-language querying and prompt-free semantic visualization. Analysis of camera-to-radar concept transferability shows stronger agreement for structures, materials, and human poses than for colors and printed text.

This paper presents a two-stage learning framework for versatile and agile humanoid locomotion that couples online reduced-order model (ROM) reference generation with whole-body tracking control. In the first stage, we train a command-conditioned whole-body reference generator using single rigid body dynamics combined with full humanoid kinematics. In the second stage, the frozen ROM policy is rolled out online to generate references for training a whole-body tracking policy on the full-order dynamics model. Learning a kino-dynamically consistent reference generator in the simplified ROM space enables us to generate references over a broad command space and over multiple locomotion strategies using a single unified framework. We demonstrate the approach on a Unitree G1 humanoid robot, achieving longitudinal velocities from -2.0 m/s to 2.8 m/s, lateral velocities up to +-1.0 m/s, angular velocities up to +-3.0 rad/s, one-legged hopping, and continuous transitions across the commands, all using one unified reference policy and one whole-body tracking policy.

In micromanipulation, erroneous hand motion limits unaided manipulation accuracy, which can affect task performance and safety in critical applications like microsurgery. Handheld robots (micromanipulators) enable motion accuracy beyond human limits by providing robotic assistance in a low-cost, compact, intuitive form factor. Particularly, active constraints are an effective form of robotic assistance, which aids the operator by restricting the robot's permissible motion based on task-specific requirements or objectives, thereby providing task guidance and enforcing safety. However, existing formulations of active constraints for handheld systems are limited, most notably because of the inability to provide guidance and enforce safety simultaneously. Implementing active constraints with optimization-based methods can address these limitations, but existing implementations are designed for free-standing robots, and neglect that handheld robots actuate relative to the operator's hand with a limited workspace. In this work, we propose an optimization-based active constraint framework for handheld micromanipulation that enables simultaneous guidance and safety enforcement. The formulation was implemented in a handheld robot and demonstrated on a medically relevant artificial task. Results show that the proposed formulation reduced safety violation rate by more than two orders of magnitude relative to baselines, while recovering the threefold accuracy loss from enforcing safety alone.

Vision-language-action (VLA) policies can learn to follow execution preferences, such as where to grasp an object or how to move around an obstacle, from demonstrations annotated with preference labels. Providing these annotations for every new task, however, is costly. We study whether learned preference control can transfer to new tasks whose demonstrations contain the desired behaviors but lack preference annotations. We find that a policy trained with preference annotations can follow preferences on a new task before adaptation but achieves limited task success. Fine-tuning on target demonstrations without these annotations improves task success while weakening preference control. Motivated by this finding, we introduce Self-Labeled Preference Transfer (SPT). SPT jointly trains a shared vision-language backbone for preference-conditioned action prediction and preference recognition on annotated source demonstrations. It then freezes the recognizer to preserve its cross-task preference recognition capability and uses it to self-label target demonstrations for policy fine-tuning. Across five types of execution preferences in simulation, the recognizer achieves 96.4% accuracy, and SPT reaches 84.8% preference following and 89.2% task success, compared with 59.4% and 88.0% for naive finetuning. We further extend to multiple preferences, achieving 93.8% preference following compared with 42.5% of baselines. In real-world experiments, SPT increases preference following from 52.5% to 92.5% relative to naive finetuning across two task transfers. These results show that SPT can support preference adaptation across tasks and objects while maintaining success, without additional manual preference annotations.

Humanoid controllers often require task-specific training despite sharing the need for balance and whole-body coordination. We present Metatron, a promptable whole-body execution layer combining a generative base policy with a programmable skill layer. The base policy receives partial goals, collectively called a prompt, that can combine timed body-part targets, velocity commands, and style cues. It produces diverse whole-body motions consistent with these goals. Training combines motion capture, synthetic recovery trajectories, and retargeted scene-interaction motions. After training, the policy is frozen. Stateful programs called skills express task intent as prompts at each control step, drawing rough geometric sketches that the policy turns into coordinated movement. Skills are ordinary code, refined by coding agents that analyze motion data and fit compact rules without task-specific policy training. On the physical Unitree G1, one frozen checkpoint supports walking, recovery, teleoperated object pickup, and sitting on furniture. Comparative evaluations show higher reaching success and push robustness than the tested reference-tracking pipelines. The same recipe produces climbing and acrobatics on a simulated character. Metatron enables coding agents to turn learned motor capabilities into new skills through code, without further policy training.

Reinforcement learning holds the promise of robots that learn manipulation skills from interaction alone, without pretrained priors or demonstration data. In practice, on-policy methods such as PPO routinely stall on long-horizon, contact-rich tasks: dense reward shaping carries the policy through stages, yet stitching them together and discovering sparsely rewarded events rarely succeeds. We trace this failure to the missing action discrimination in state-value credit assignment: a state-value critic judges progress only by episode outcomes, so when the policy converges to a partial solution all outcomes coincide, and improvement stops because no action can prove better than another. We therefore extend PPO with an action conditioned value critic which, can compare the executed action with alternatives sampled at the same state and aid the learning signal. Combined with Generalized Advantage Estimation and a bound on exploration noise, we show how the extended algorithm escapes the plateau and completes precise peg insertion within a training budget where PPO with identical exploration fails.

We introduce EARS-Info, an information-aware extension of Event-Driven Ant Reallocation Search for multi-robot exploration on OpenStreetMap-derived grids. It combines marginal sensing gain, feasible relocation, return-aware routing, sensing reservations, and cached shortest paths. Across 168,000 episodes comparing 30 implementations on eight city maps, EARS-Info achieves the highest mean operational score in every evaluation panel. With eight robots and 3% communication dropout, it improves the score over EARS-Safe by 5.43%, increases weighted discovery from 88.88% to 98.46%, and reduces mean energy use by 2.62%. Score differences are positive in all eight cities ($p=0.0078125$; descriptive 95% city-bootstrap interval $[0.0159,0.0694]$), with no simulated incidents across 5,600 EARS-Info episodes. These results support the complete controller; individual component advantages remain uncertain. Generalization is limited by reused geography, one training replicate, coarse maps, and adapted comparison methods.

Recent advances in monocular depth estimation have demonstrated highly accurate 3D reconstructions within in-distribution terrestrial domains. This study investigates the efficacy of several state-of-the-art approaches to reconstruct out-of-distribution lunar environments using the NASA Ames Polar Optical Lunar Analog Reconstruction (POLAR) stereo dataset and Apollo 17 surface imagery. By leveraging these datasets, we establish a robust analog-to-field validation pipeline that evaluates quantitative performance against ground-truth physical analogs and qualitative performance against historical field data. Our results indicate that select state-of-the-art techniques can successfully generalize to produce high-fidelity relative 3D reconstructions of the unstructured lunar surface.

Robot-learning studies often compare position and velocity actions as though each label identified a single experimental treatment, yet the label alone does not determine what the robot executes. Before affecting robot motion, bounded policy output passes through filtering, physical scaling, reference construction, feedback, timing, and effort limits. We study how these omitted properties affect attribution under one shared learner and training budget. Across three contact-rich Factory tasks (PegInsert, GearMesh, and NutThread), we compare five joint-space interfaces, with ten independently trained policies per condition. We first remain within the same position-feedback law and vary the executed interface: substituting only the physical action scale as the directly modified component changes sample-median evaluation success by 0.645-1.000 across tasks, while removing an exponential moving average produces task-dependent changes. We then move to a cross-law comparison and choose a velocity scale that exactly matches the position interface's local raw-torque command gain. This calibration raises the sample median on GearMesh and NutThread, with no median change on PegInsert. Yet clear residual sample-median gaps remain on PegInsert and GearMesh, while the NutThread interval includes zero. Thus, even this explicit equality leaves feedback and held-action dynamics unmatched and does not establish position-velocity trained-outcome equivalence. Action-interface comparisons should therefore report the executed pathway and state the physical quantity they match.

We continue the study of minimum-time escape of an omnidirectional evader from the forward conic field of view of a Reeds--Shepp pursuer, now for a faster evader, and solve the game for every speed ratio $ν>1$ and every half-angle $0<α<π/2$: the escape time is characterized by explicit characteristics and the selection of roots of a scalar equation, and it is the value of the game in nonanticipative strategies at every state. Above the end $E$ of the evader's dispersal line on the front axis, the evader chases the reversing pursuer along a universal line fed by tributaries. The characteristic from the boundary of the usable part is an involute of a circle of radius $ν$ ending at a cusp, and two explicit thresholds decide whether it reaches the axis before the cusp and whether the jump of the escape time across it ends before the cusp or runs to it; in the latter case primaries fold back next to the barrier. The pursuer's optimal strategy is the relay of the slower case. The evader follows the tributaries and switches once to a primary, or commits in advance to a heading certified by the scalar equation of the primaries. Level-set and semi-Lagrangian computations are compared with the value.

Future planetary exploration missions require teams of robots to collaboratively map unknown environments under severe communication limits, constrained onboard computation, and no central infrastructure. In this field report, we present SKiD-SLAM, a distributed multi-robot LiDAR simultaneous localization and mapping (SLAM) system for such environments, and report its deployment across two mission concepts: cave exploration and scientific surface exploration. We survey the communication and computation budgets of actual planetary missions and select modules that fit these budgets for inter-robot place recognition, relative pose estimation, and distributed trajectory fusion. Before field deployment, we verify the assembled system on public datasets resembling planetary environments and in a simulator built on NASA JPL's Open Source Rover. We then validate SKiD-SLAM through field deployments in South Korea: a team of wheeled rovers explores the Majang lava-tube cave on Jeju Island, and quadruped robots carry out a crater-and-rock science scenario on a planetary analog terrain at the Korea Institute of Civil Engineering and Building Technology (KICT) in Yeoncheon. Across both deployments, SKiD-SLAM reliably performed distributed multi-robot LiDAR SLAM within a 250 kbps inter-robot communication budget, demonstrating its feasibility under challenging field conditions. Supplementary materials, including code and field data, are available at https://sparolab.github.io/research/skid_slam/.

We present an experimental framework for determining sufficient communication in robot teams running fixed, independent Behaviour Trees with private blackboards. Within a declared policy family, we vary message content and admission frequency and measure traffic in recipient-weighted bits per robot decision. A dense sweep identifies candidate policies for comparison with full-state communication. We retest selected candidates and the reference on fresh, paired scenarios, using one-sided simultaneous confidence bounds to assess whether success remains within five percentage points of the full-state reference. The bounds also assess whether tested cheaper policies can be ruled out as insufficient. Depending on the evidence, the framework reports an exact minimum, a positive bracket, a sufficient upper bound, or an unresolved boundary. Experiments with eight content policies across four warehouse conditions identify zero communication as the exact minimum under low contention. In the scaled condition, the no-release policy at gate $q=0.992222$ is confirmed sufficient, achieving 96.33% success at an estimated rate of 738.55 bits per robot decision, or 0.544% of full-state traffic. The positive minimum remains unresolved. These claims apply only to the declared policy family, sampled gate values and operating conditions.

Behavior Cloning has demonstrated promising generalization capabilities for robotic manipulation in terrestrial environments. However, marine visual distortions and dynamic currents create a severe domain shift that remains a key bottleneck in adapting generalist architectures to underwater field manipulation. To address this, 1616 episodes of underwater manipulation data were collected in calm and turbulent waters and were used to train 14 generalist Vision-Language-Action (VLA) models (DynamicVLA) with varying proprioception observation configurations. Across three tasks on 1260 gradable on-hardware rollouts in a pool, 10 models outperformed the baseline, with the highest-performing model achieving a 93.3% overall success rate, including partial successes. Results are also introduced from a case study evaluating the performance of a single-task Action Chunking Transformer (ACT) and a generalist VLA (SmolVLA) policy for two tasks in marine archaeology and conservation: grasping and mooring. Experiments are performed on an ROV on and around the WWII wreck Coralita X127 and at Manoel Island Marina, Malta. Across 748 collected episodes and 508 gradable on-hardware rollouts, the ACT policy substantially outperformed SmolVLA (47.9% vs. 7.9% pooled success rate), earning a 66.7% success rate on the most successful task at the marine archaeological site. To enable further contributions to embodied intelligence in marine field robotics, the BLUEMBODIED KIT used to collect data, train models, and deploy policies on the ROV has been open sourced.

Robot imitation learning uses demonstrations from operators of varying skill, but selecting training data remains challenging. Existing methods score demonstrations using smoothness, diversity, or estimated influence on policy performance, but differences in selection and evaluation protocols make it difficult to determine whether they outperform simple baselines. We introduce \textbf{CuraEval}, a framework combining mathematical analysis of demonstration scores with controlled comparisons against fixed size-matched random subsets and full-data training on mixed-skill robomimic data using Diffusion Policy and BC-RNN. We show that the evaluated PSD score depends only on episode length and squared position magnitudes, and produces nearly the same ranking as episode length on \textit{square} ($ρ=0.999$). With Diffusion Policy, selecting the shortest 50\% of demonstrations improves \textit{square} success from 70.7\% under random selection to 76.5\%. Compared with random selection, our adapted geometric-entropy LOO selector reduces success at 25\% retention from 94.7\% to 74.0\% on \textit{can} and achieves 41.7\% on \textit{square}, below the mean success rates of all three independently sampled random subsets (59.8--62.5\%). On the same 3 \textit{square} Diffusion Policy training seeds, length-based selection at 75\% retention improves success over full-data training from 76.7\% to 86.7\% under the original protocol, but lowers it from 80.5\% to 79.5\% after checkpoint reporting, rollout count, and evaluation seed change jointly. No evaluated curated condition significantly outperforms full-data training after Holm correction, although confidence intervals still permit useful gains. Code is available at https://github.com/jiuyixu25/CuraEval.

Mobile manipulator robots rarely reposition their base during manipulation, because doing so requires real-time knowledge of how well the arm can reach its surroundings among obstacles. Reachability maps (RMs) encode this knowledge, but existing methods are slow, coarse (typically 5 cm), or ignore real obstacles. This article asks whether a neural network can predict an obstacle-aware RM at 2 cm resolution from a live depth camera, on embedded hardware, fast enough for online use. We present HERON, a 3D U-Net trained on RMs computed with the GPU-accelerated inverse kinematics of cuRobo, first on synthetic scenes and then on real depth data, and deployed on a Jetson AGX Orin with TensorRT and INT8 quantization. HERON predicts the RM of a 2.56 m cube around the arm in 115 ms on average, for a measured system rate of 8.57 Hz, close to the 10 Hz update rate of the obstacle map, whereas computing the same map with cuRobo takes about 90 min. Against the cuRobo ground truth, 95% of the voxels lie within 1 class out of 51 with the FP32 model and within 3 classes with the INT8 model, with a Pearson correlation of 0.99 and a quadratic weighted kappa above 0.9. Accuracy degrades mainly at the workspace boundaries.

We demonstrate the first zero-shot sim-to-real transfer for a surgeon-robot collaboration task in open surgery: retractor take-over. A robot arm takes a retractor from the surgeon's hand and preserves retraction at handoff, using a policy trained only on simulated demonstrations, without human demonstrations or real-world fine-tuning. Imitation learning from real demonstrations is limited by collection and labeling costs, and by performance degradation when the environment changes. Our framework, SimRetract, generates simulated demonstrations designed for transfer: the policy observes masked instrument images and grasp geometry, learns to recover from off-nominal states, and acts with margin for execution drift. Deployed zero-shot, it grasps the retractor in 90% of 50 surgeon-defined phantom presentations and preserves retraction at handoff in 78%. Results on unseen retractors suggest potential for generalization to instruments with similar grasp geometry. Across ten policy configurations and 1,200 simulated and real trials, we localize the sim-to-real gap to components: in simulation, grasp rates across ablations span 80-96%, while on the real robot removing the instrument mask alone drops performance to 22% grasped and 12% held. Transfer-oriented design thus makes zero-shot sim-to-real transfer feasible for a surgical assistance task without human demonstrations. Code is available here: https://anonymous.4open.science/r/simretract-58D2.

Vision-and-Language Navigation in Continuous Environments (VLN-CE) requires instruction-consistent planning and reliable execution under collision constraints. Yet many hierarchical approaches optimize the planner while keeping execution fixed. We introduce MATCH, a framework that jointly learns a language-conditioned planner and a goal-conditioned primitive-action executor. An execution-aware topological graph represents start-dependent execution options and integrates executor value estimates with realized outcome memory, enabling the planner to select execution requests and refine their targets through bounded continuous offsets. A transition-aware dense macro reward balances ordered progress along the reference route against execution residual, encouraging executable requests without substituting local success for instruction following. A staged curriculum culminates in synchronized on-policy updates, with the executor learning on planner-generated requests and the planner adapting its requests to the executor's current behavior. With 0.5B parameters, MATCH achieves state-of-the-art performance on the R2R-CE and RxR-CE benchmarks. Preliminary real-robot deployments further demonstrate its practical potential.

Embodied systems can update policies, world models, skills, and executable programs from deployment experience, allowing one interaction to benefit later tasks. When a system also revises and reuses its data-acquisition, feedback, or learning procedures, experience changes the process that produces capability, introducing recursive self-improvement. This paper surveys research in physical AI through the connection between persistent capability and its improvement process. We first explain the functional distinction between learner updates and improvers, then compare policy learning, imagined experience, memory and skill accumulation, and executable harnesses. These approaches connect to revisions of curricula, rewards, and research programs in different ways. Contact validity, recovery, retention, and resource costs determine whether the resulting gains accumulate during continued deployment. An L0-L3 rubric distinguishes task-local adaptation, persistent capability gain, inherited process revision, and verified improver gain. Representative benchmarks and within-study contrasts relate these distinctions to existing results. We further propose common-data and interactive comparisons that evaluate reference and revised improvers from shared learner starts, separating the use of experience from its acquisition. This perspective connects the diverse methods, physical operating conditions, and evaluation evidence, underpinning embodied RSI. A comprehensive list of studies in this survey is available at https://github.com/nkd-lkz/physical-ai-rsi-survey.

Compositional diffusion planners have demonstrated remarkable success on long-horizon diffusion planning by sampling short trajectories, namely chunks, and composing them into a long trajectory. However, averaging overlapping chunks can induce the mode-averaging problem, resulting in discontinuous trajectories. Several works aim to address this problem, whereas a non-negligible portion of plans remains infeasible. In this work, we propose Selective Compositional Diffuser (SCD), handling this problem by measuring reachability. To this end, we present a self-supervised learning method for the reachability classifier. This is trained on offline data and generated chunks jointly to address potential distributional mismatch, while reducing the bias inherited from the generated data. Based on this, we construct a graph whose edge costs represent whether consecutive chunks are reachable, and traverse the graph to find effective yet efficient routes. Lastly, we generate bridging trajectories and refine the composed trajectory for smooth transitions. We evaluate our method on OGBench stitching tasks. It practically reduces incompatible mode compositions and the mode-averaging problem compared to the studied baselines. Across tasks, SCD improves success rates, reaching state-of-the-art results on HumanoidMaze-Large and Giant tasks.

Zero-shot vision-and-language navigation (VLN) requires an agent to use past observations without navigation-specific training. We study what information an 8B-scale navigation system should retain and how it should be organized for subsequent decisions. Our agent combines instruction progress, hierarchically compressed observation history anchored to poses, and accumulated geometry in complementary task, temporal, and spatial memories. An 8B observer summarizes observations, and an 8B reasoner uses the memories to select waypoint intents or request additional views, a geometric controller executes the selected targets. A semantic information perspective motivates the distinction between retaining evidence, making it accessible to the reasoner, and acquiring further observations. On R2R-CE val-unseen, the complete system achieves the highest among the evaluated Qwen3-based zero-shot baselines. Removing temporal memory reduces SR and sharply increases average path length from 16.9 to 39.1m, removing the spatial-map and planning component reduces SR as well. Real-world experiments on a Unitree Go2 show that our zero-shot method achieves higher success rates than the post-trained baseline in both corridor and office environments, suggesting better transfer to the evaluated real-world settings. As a complementary evaluation, LAW2 uses paired instruction interventions to examine behavioral responses, instruction following ability beyond endpoint success. These results support preserving task progress, observation history, and geometry in structured memory for zero-shot navigation with smaller open models.

Reliable box manipulation requires successful placement without damage, which depends on maintaining sufficient contact force without excessive loading. We hypothesize that per-finger force feedback improves manipulation reliability relative to open-loop finger commands. To this end, we present Finger External-Force Estimation from Proprioceptive SignaLs (FEEL), an uncertainty-aware residual method that estimates per-finger external force magnitude from proprioceptive histories. Our method combines a calibrated virtual-spring estimate with a residual distribution predicted by a Mixture Density Network (MDN) and requires no dedicated contact sensors during deployment. The estimated force is used for continuous per-finger adaptation, while predictive uncertainty is evaluated as an empirical signal of ambiguity in proprioceptive force estimation. Across ten fingers and contact forces spanning 0-20 N, our method achieves a mean RMSE of 0.251 N, reducing the error by 92.8% relative to the virtual spring and by 9.4% relative to an end-to-end MDN. In box pick-and-place, force adaptation maintains task success at 10/10 and increases both damage-free outcomes and reliability from 7/10 to 10/10. The project page is available at https://finger-force-estimation.pages.dev/

Discretizing continuous actions into discrete skills using methods like VQ-VAE has emerged as a powerful paradigm for robotic manipulation. However, quantization errors yield a suboptimal distribution for the skill prior, degrading its performance. While reinforcement learning offers a path for refinement, its direct application faces unstable encoder updates and a granularity dilemma in importance sampling. To address these challenges, we introduce Cascaded Skills Optimization (CSO), a two-stage post-training framework. First, Rejection-Sampling Supervised Fine-tuning aligns the model's observation-to-skill mapping with the distribution of successful online trajectories via supervised fine-tuning. Second, with the observation encoder frozen, Skills Policy Optimization computes an independent, clipped importance ratio for each complete skill, enabling stable and efficient updates. With a lightweight QueST-style policy, CSO raises overall LIBERO success from 80.1% to 97.5%, improves performance on RoboTwin2.0, and achieves 61.0% success across five physical single-arm tasks. Additional comparisons with FAST, discrete BEAST, and fixed-buffer IQL further validate the effectiveness of our two-stage post-training strategy.

Quadrupedal manipulators can extend their workspace through whole-body interaction, but task-planning agents' capabilities remain limited by their available controllers. We present ELEVATE, a framework that uses persistent failures during task-program refinement to identify and acquire missing whole-body control capabilities. When these failures suggest a missing physical capability, a coding agent formulates and iteratively refines a reinforcement-learning problem, specifying behavior-specific rewards, curricula, and termination conditions. Validated controllers are added to the capability library and composed with existing controllers to retry the task. Starting from a single pretrained loco-manipulation controller, ELEVATE autonomously acquires controllers for wall-supported elevation, supported manipulation, and wall descent across a sequence of seven simulation tasks. These controllers are reused in subsequent reaching and object-manipulation tasks. On the five tasks beyond the initial controller's workspace, ELEVATE achieves mean success rates of 98.4%-100% across five independent runs, while coding-agent baselines with a fixed controller library fail to complete these tasks. The synthesized task programs, together with the acquired controllers, also transfer to a physical quadrupedal manipulator without real-robot training or fine-tuning, achieving 18/20 successes in rack retrieval and 16/20 in floor-to-rack placement. Videos and demonstrations are available at https://agentic-quad-elevate.github.io/public/.

When a person declined our robot's invitation to play, Alice turned her shoulder away and said "No worries, it's all good." Only after a brief pause did she drive off. The words and gestures were plausible but their order changed the message. In another episode, Alice informed a host that his visitor had arrived---while the visitor stood right beside him. Both came from a three-year ROS 2 HRI consortium that integrated large language and vision models as they emerged. Today's foundation models give robots broader capabilities to perceive their surroundings, reason about situations and generate behavior. Yet capability alone does not ensure that a changed situation shapes the action that reaches a person. Evidence may never be acquired, or a purpose or constraint may fail to survive planning and execution. A system may even report Success when an action no longer serves its intended purpose. Drawing on the consortium's engineering experience, we develop five design lessons about attention, knowledge, commitments, embodied timing and action authority. We then introduce TRACE, a lightweight review that changes one interaction premise, follows it to a decision or enforcement boundary, and produces an actionable architecture finding. A bounded analysis of a published LLM-generated behavior-tree system illustrates how an available clarification mechanism can go unused. The engineering lesson is to preserve accountability boundaries, not the computational pipelines that once exposed them.

To solve problems in 3D space rather than merely describe images, multimodal large language models (MLLMs) need 3D spatiotemporal reasoning: reconstructing 3D geometry from occluded 2D views, anticipating physical dynamics, and executing precise, closed-loop actions. Prior studies test passive VQA on static images or symbolic planning, requiring no execution or supplying privileged coordinates while ignoring collision, statics, and timing. We introduce SpaceAct, an interactive benchmark evaluating 3D spatiotemporal reasoning via closed-loop physical action and active perception, with SAPA, a Spatiotemporal Active-Probing Agent framework. Built on interlocking bricks and mechanical puzzles, it has 108 verified instances across 36 task families in six categories, without privileged coordinates, and diagnostic modes (+plan, +state, no-camera, blind, L7 levers). In brick-bridging or aperture-threading, models rotate the camera to resolve occlusion, read in-scene rulers, output 6-DoF poses under swept-collision and statics checks, and self-correct via split-screen rollback feedback (dynamic tasks freeze time during deliberation and give five strobe frames). Across 10 frontier MLLMs on the 44 hardest instances, scores span 4.8-79.1 (seven solve at most 27.3% of 6-DoF tasks); for the top three, a reference plan (+plan) closes 44% of the gap while leaving collisions at 5.2 per episode, whereas adding exact 3D state (+state) closes 90% and cuts collisions to 0.5; for four others, +plan lifts the mean score from 10.9 to 35.3 and +state to 81.8; and 46% of failures are collision or observation loops. SpaceAct and SAPA advance evaluation to closed-loop physical interaction, showing that even with a valid plan, MLLMs remain bottlenecked by 3D visual grounding and action refinement, while offering a testbed for active multi-view exploration and explicit 3D grounding.

Robotic grasping relies on large manually annotated datasets, which are costly to create. We introduce a straight-skeleton-based algorithm for generating two-finger grasp keypoints and identifying regions suitable for stable, feasible grasps. Our automated pipeline produces consistent grasp annotations without human intervention. We also propose the Skeleton-based Generative Grasp Convolutional Neural Network (SkelGG-CNN), a 4-DoF architecture that learns geometric grasp features from skeleton-based guidance. To extend auto-labeling to 6-DoF grasping, we develop a generalized quasi-3D skeletonization algorithm that operates directly on point clouds. Using automatically generated 4-DoF labels, SkelGG-CNN achieves accuracies of 95.42%, 96.55%, and 97.1% on Jacquard V1, Jacquard V2, and Cornell, respectively. On the GraspNet novel split, 6-DoF auto-labeling achieves an average precision of 18.33, comparable to state-of-the-art model-based approaches. In simulation, the 6-DoF method achieves success rates of 98.4% on Dex-Net and 89.1% on EGAD!. Real-world experiments with a Delta parallel robot show success rates of 97.4% on single household objects and 95.0% in cluttered scenes, outperforming state-of-the-art methods without domain adaptation. The 4-DoF pipeline adapts across environments and produces annotations approaching human-level precision, enabling efficient training of lightweight models on small datasets. The generalized 6-DoF method provides scalable, geometry-aware annotations for cluttered 3D scenes.

Combining wheeled and legged locomotion within a single robotic platform offers the potential to harness the advantages of both systems. The development of such hybrid systems remains an active area of research. While legged robots exhibit remarkable adaptability for traversing uneven and rugged terrains, their relatively low speed limits their practicality in indoor environments. Conversely, wheeled robots demonstrate higher efficiency on flat surfaces but often encounter difficulties when navigating obstacles such as steps. This paper introduces OmniQuad, a novel hybrid robot that integrates mecanum wheels with legged locomotion to exploit the benefits of both modalities. Mecanum wheels provide omnidirectional mobility, enhancing the robot's manoeuvrability by enabling agile movement in confined spaces and the ability to maintain a consistent orientation during task execution. In this work, we first present the design of a custom wheeled-foot featuring mecanum wheels. Subsequently, we describe the complete robot assembly, comprising these wheeled feet and EM-Act actuators. Finally, we detail the control scheme developed to govern the OmniQuad and evaluate its performance through experimental trials conducted in both indoor and outdoor environments.

Joint-embedding predictive architectures (JEPAs) learn to predict representations rather than pixels, offering a compact way to model how visual scenes evolve. We ask whether this predictive objective improves closed-loop control when a vision-language-action (VLA) policy already uses V-JEPA features. We call our JEPA-conditioned flow policy family SaiVLA-JEPA. It combines a vision-language model with a diffusion transformer action expert and receives features from a frozen V-JEPA encoder. Two lightweight auxiliary heads predict future representations during training and are omitted at inference, leaving the action sampler unchanged. We compare policies without JEPA features, with current JEPA conditioning, and with additional future supervision on 10,030 paired LIBERO-Plus episodes. Current JEPA conditioning reaches about 82.5% success versus 71.6% without JEPA, while future supervision provides no detectable aggregate gain at the evaluated training budget. We separately evaluate SaiVLA-JEPA-TCA, which adapts latent context, and SaiVLA-JEPA-TPA, which adapts selected parameters. The context-retry gain is a system-level result under a larger attempt budget; parameter adaptation shows no consistent advantage over the frozen policy and target controls. These results motivate assessing latent-prediction objectives through paired closed-loop rollouts rather than latent loss alone.

The growing volume of end-of-life products has intensified the need for automated disassembly, yet hook-type connections remain challengin: their orientations vary, they deform under load, and their geometry is hidden, so vision alone cannot identify a feasible non-destructive disassembly direction. This work proposes a constraint-aware reinforcement learning framework that infers feasible extraction directions online from contact forces. The policy is trained in a robot-agnostic, object-centric simulation on a simplified peg-in-hole pair and deployed zero-shot on a physical robot, where it transfers to real connections with unseen, multi-stage constraint structures. Because instantaneous contact forces are noisy and locally ambiguous, the policy is coupled with a particle filter belief that provides the policy with temporally integrated directional information and reduces redundant exploratory motion. The approach is validated on four real-world tasks, from peg-in-hole pairs to the inner propeller and front cover of an end-of-life air-conditioning unit. The full method achieves success rates of 100%, 100%, 90%, and 75% with the highest path efficiency on every task, and raises the success rate on the air-conditioning cover from 45% to 75% relative to a policy trained without the belief. These results indicate that force-driven reinforcement learning can enable disassembly without product-specific models or demonstrations.

Smart wheelchairs are robotic devices that gather information from the surroundings and allow people to safely drive through various environments. They usually rely on the interpretation of high-level user's intentions, typically decoded through human-robot interface (HRI) or brain-machine interface (BMI). However, navigating in an unknown environment and accurately understanding user commands remain significant challenges. Misinterpretation of commands or the environment can lead to user frustration, distrust of the system, or even unsafe situations. In this paper, we present a novel Multi-Layer Navigation System (MLNS) for a smart wheelchair that enables navigation in unknown environments while interpreting and safely contextualizing user's intentions even in the case of limited and faulty commands. The stack relies on a Virtual Sensors-Driven Potential Fields (ViSPF) method for obstacle detection that accounts for the wheelchair's actual size, a Model Predictive Control (MPC) module for trajectory generation, and an open-vocabulary computer vision algorithm for target identification. Finally, a system based on Priority Queue (PQ) was implemented for target selection. We validated our system on a real smart wheelchair with experiments involving 10 participants without disabilities, including scenarios with faulty input interfaces. The results demonstrate the robustness and effectiveness of our system in real-world environments, proving its potential for practical applications in assistive technologies.

Reliable perception is essential to intelligent transportation systems, yet deep neural networks can silently miss safety-critical objects while continuing to produce apparently plausible outputs. Existing run-time monitors for 3D object detection generally assess each frame independently and therefore cannot exploit the temporal evolution of occlusion, point sparsity, and scene complexity. We present TRIM, a lightweight temporal monitor that detects missed-object failures without modifying or retraining the underlying detector. TRIM summarises intermediate activations using compact channel-wise statistics and processes short sequences of these summaries with a temporal convolutional network. Experiments on NuScenes show that temporal monitoring improves AUROC from 0.686 to 0.704 and reduces the false alarm rate at 95% recall from 0.818 to 0.759 compared with single-frame monitoring. Early and intermediate backbone features provide the strongest monitoring signal, while the temporal inference head requires only 0.159 ms on a CPU.

Future planetary rovers will traverse steep, deformable terrain where slip, sinkage, and terrain deformation can threaten mobility. Although simulation can capture some wheel-soil effects, unmodeled deformable-terrain phenomena create sim-to-real gaps requiring deployment-time adaptation. Yet, rover operations leave little room for extensive trial-and-error. We introduce TERRA (Targeted Exploration for Reward-Relevant Adaptation), an online adaptation framework that efficiently seeks deployment data exposing task-relevant model errors and corrects them. TERRA adapts a simulation-trained world model with a reward-weighted prediction loss and trains a loss predictor to estimate where future performance-relevant errors are likely to occur, using this signal to guide exploration during MPC. In simulation, TERRA reaches approximately 90% slope-climbing success within five trials under severe out-of-distribution pretraining conditions, outperforming exploration baselines in reward, completion time, and energy to first success. On the ERNEST rover, TERRA recovers coordinated active-suspension locomotion on a deformable Mars-simulant slope within three to five real-world trials. Project website: https://terra-explore.github.io

Post-training quantization (PTQ) reduces the memory footprint of vision-language-action (VLA) models, but lowering bit widths while maintaining task performance remains challenging. Existing work has used sensitivity estimates from calibration data to assign bit widths to different model components. However, using task success directly to guide mixed-precision allocation is underexplored because closed-loop evaluations are costly and the effect of quantizing one component depends on the precision of other components. To this end, we propose \textbf{QbD-VLA}, a task-guided mixed-precision quantization method that uses task success from short closed-loop evaluations to assign weight and activation bit widths while accounting for interactions between model components. Experiments covering four VLA models and all four LIBERO suites show strong task performance at average precisions of approximately 4 bits for weights and fewer than 6 bits for activations. On SIMPLER, the gain over uniform quantization is small. On a Franka Research 3 robot arm, simulation-selected allocations show no detected success-rate loss on X-VLA but fail on $π_{0.5}$. These findings support using task success to guide mixed-precision VLA quantization, while highlighting the need for real-world evaluation before deployment. Code is available at \href{https://github.com/jiuyixu25/QbD-VLA}{\texttt{github.com/jiuyixu25/QbD-VLA}}.

Autonomous navigation in off-road terrain is commonly achieved using hierarchical control and planning architectures that rely on explicit cost maps and one-way information flow from a high-level planner to a low-level controller, decoupling planning from the true execution cost experienced during control. We propose a hierarchical planning and control framework in which information flows bidirectionally, allowing the low-level controller to guide trajectory selection at the planning level. The highest level of the hierarchy is a Model Predictive Path Integral (MPPI) control-style planner that determines optimal trajectories and passes these waypoints to a pretrained low-level reinforcement learning (RL) controller. Rather than drawing from an explicit cost map, our MPPI planner leverages the value function of the RL controller as an implicit cost representation. In this formulation, MPPI serves as a trajectory exploration mechanism, while the low-level controller's value function guides the search. To account for the distribution shift seen by the value function between training and evaluation, we propose an online policy evaluation process that adapts the value function to unseen driving experiences, and we demonstrate this adaptation in unseen off-road simulation environments. We evaluate our approach in a high-fidelity off-road driving simulator across a range of challenging terrain conditions and compare its performance against a baseline MPPI controller. The results demonstrate that our method achieves safer and more effective navigation to goal locations while eliminating the need for explicit cost maps and enabling tighter coupling between planning and control.

Embodied instruction following in interactive environments requires agents to interpret natural language commands and complete long-horizon tasks. Existing planners either rely on large labeled trajectories, or use few-shot and zero-shot LLM prompting that produces a whole plan in a single serial reasoning process. These approaches depend on linguistic common sense, commit to a plan from the initial instruction, and only update after failures, which weakens robustness and state awareness. We present AtomTree, a hierarchical framework for state-aware task planning with LLMs. AtomTree incrementally builds a tree of subgoals and actions, where the LLM decides at each step, from the current symbolic state, whether to further decompose or to output a grounded low-level action. A lightweight memory module summarizes past outcomes for reuse in later decisions. When a low-level action stalls, AtomTree triggers hierarchical backtracking replanning to return to the parent node and regenerate the subtree plan from the latest state, preventing continued drift along an incorrect trajectory. In zero-shot evaluation on the ALFRED benchmark, AtomTree outperforms prior LLM-based planners. Project page: https://zhaihaotian.github.io/atomtree-website/.

Tracked vehicles are widely used in off-road applications, but their conventional motion models often rely on no-slip assumptions that become inaccurate on low-adherence surfaces. Rather than treating inertial effects arising from slip as a disturbance to be minimized, this work investigates how inertial effects arising from a moving mass can be deliberately exploited for steering on low-friction surfaces. A dynamic model for an articulated tracked vehicle with a passive trailer is developed, allowing inertial effects from trailer motion to be captured. The approach is experimentally evaluated on ice, where our dynamic model outperforms the kinematic baseline. We also propose a mixed dynamic-kinematic model to adapt predictions to both surface and situation by exploiting predictions from both methods. Finally, we demonstrate inertial steering on a mono-track vehicle without a conventional steering mechanism, showing that controlled movements of an internal mass can generate turning motion on ice. The resulting behavior can be predicted using the proposed dynamic model by representing the internal mass as a frictionless trailer, revealing the analogous dynamics between the two configurations. These results demonstrate the feasibility of inertial steering as an alternative steering mechanism for tracked vehicles operating in low-adherence environments.

Quadrotor control under disturbances such as wind gusts and payload variations remains challenging. First-order sliding mode control (SMC) and matched/unmatched uncertainty distinction are classical approaches; however, neither tells a practitioner which loops in a cascaded controller should be made robust to sustain disturbances, how to tune those loops, nor what the additional costs are on an embedded processor. Applying robust control uniformly across all axes addresses disturbances indiscriminately, adding tuning effort on axes where no disturbance acts. On inherently underactuated vehicles, disturbances along some axes cannot be opposed directly at all. We propose a principled methodology for deciding where robust control belongs in a cascaded architecture, with all control running onboard the vehicle. We recover the acting disturbance from PID hover flights under disturbance by inverting the equations of motion, use the recovered signal to identify and quantify the excited axes, and deploy robust control where it counts. Experiments on a Crazyflie 2.1 quadrotor under slung-payload and oblique-wind disturbances show tracking improvements concentrated on the axes the rule selects, and gain sweeps confirm that the switching gain reduces error on matched axes but has no effect on an excited unmatched axis, as the analysis predicts.

Reliable obstacle avoidance is essential for safe quadrotor flight in unknown environments, yet conventional frame-based cameras remain challenged by motion blur, limited dynamic range, and latency under fast motion. Event cameras, with sub-millisecond latency, high dynamic range, and negligible motion blur, offer a promising alternative, but leveraging them for closed-loop control remains difficult. Existing approaches either adopt cascaded architectures that first estimate depth and then map it to control commands, introducing error propagation, or achieve end-to-end control through costly offline pretraining before online fine-tuning can even begin. We address these challenges with a CUDA-based event simulator that exploits GPU parallelism to render events efficiently within the training loop, integrated into a differentiable simulation pipeline that enables direct gradient backpropagation through the differentiable physics model, improving sample efficiency. This allows us to train an end-to-end policy that maps event frames directly to control commands, without intermediate depth estimation. We validate our approach through zero-shot real-world deployment on a quadrotor, achieving robust obstacle avoidance in both odometry-assisted and odometry-free settings, including under low-light conditions where conventional cameras typically fail.

Anticipating the motion of nearby objects is critical for safe and efficient robot navigation. Without a proper understanding of what may move, a robot may plan overly conservative, inefficient trajectories to guarantee safety, or it may be overly risky, resulting in unacceptable collisions. While tracking and predicting the motion of known objects has seen significant progress, little research addresses how a robot should adapt online to novel objects, including anticipating how one object may move another (e.g., a person operating a forklift to move a pallet). To enable this capability, we propose a meta-learned graph neural network that leverages a semantic, geometric, and relational history of objects to accurately predict, from a few online observations, when objects will begin to move in new environments. We demonstrate that our method, which is trained only in simulation, accurately anticipates the motion of objects unseen at training time, transfers to hardware without any fine-tuning, and outperforms a vision-language model and a class-frequency baseline in both simulation and hardware experiments. Our method's accurate prediction enables collision-free traversals in a closed-loop navigation simulation, while remaining within four seconds of the fastest, collision-prone planner.

Radio-frequency (RF) sensing enables robot ego-motion estimation in darkness, smoke, and fog, yet RF odometry remains challenging because multipath and moving reflectors make cross-frame matching highly ambiguous. Existing methods often rely on auxiliary inertial sensors or ground-truth pose supervision. We present RFWalk, a self-supervised, RF-only odometry framework that models dense correspondences across range-azimuth frames as random walks on a space-time graph and learns them from unlabeled sequences through temporal cycle consistency. To rule out correspondences that are cycle-consistent but physically implausible, we introduce a Doppler consistency loss that constrains correspondence-induced motion using radial velocity measured by the RF sensor itself. Ego-motion is then recovered through robust $\mathrm{SE}(2)$ registration that suppresses uncertain correspondences and those arising from independently moving objects. Without pose labels or auxiliary sensors, RFWalk remains competitive with supervised baselines in static scenes and reduces absolute trajectory error by 58% in dynamic scenes with moving pedestrians.

Catching arbitrary objects requires generalization to novel projectiles and real-time adaptation to unpredictable contact dynamics under noisy visual observations. This paper presents Catch It All!, a sim-to-real catching framework for humanoid robots that generalizes and adapts to unseen objects with variable physical properties for underhand and overhand catches, using only one egocentric camera. We learn policies which generalize to unseen objects with varying geometry, deformability, elasticity, and physical properties. Effective grasps which adapt real-time to object poses and shapes are learned through careful grasp reward design. We provide safety-aware curricula which enable safe, high-frequency action execution near hardware limits, including wrist torque limits as low as 5 Nm. We validate our system by deploying zero-shot on a Unitree G1. With only one egocentric camera, Catch It All! successfully performs underhand and overhand catches on 9 unseen objects not seen during training, including low-rebound plushes, semi-deformable toys and chip bags, and rigid-elastic objects with high rebound on contact.

Modern robot learning systems feed on high-quality demonstration data, and their performance hinges on coverage. Without good coverage, small errors compound and the policy can quickly go into out-of-distribution states from which it cannot recover. Yet there is no principled way of collecting high-coverage data, and data collection often relies on heuristic instructions to human operators. Purely online reinforcement learning (RL) explores at the cost of safety and sample efficiency, while DAgger-style interventions require the policy to fail by chance before a human can correct it. In this paper, we ask instead whether we can modify how the robot behaves during data collection to collect more useful data. We propose Data Inoculation (DI), a shared-control framework in which a human provide closed-loop teleoperation commands while a residual policy perturbs these commands. The residual policy is trained online, on the growing dataset, with the same reward-maximizing objective later used for policy extraction, so its perturbations steer data collection toward the deviations a policy trained on the data would make. These perturbations expose the robot to a wider range of states, while the human's continuous corrections provide recovery demonstrations, "inoculating" the dataset against the learner's future mistakes while the human remains in control. On three challenging real-world robotic tasks, DI consistently improves or matches the success rate of policies trained with both behavioral cloning and offline RL, showing DI as an effective data collection strategy for improving downstream policy performance.