The paper proposes a flow control method for VLA models that steers robot actions in real time through generic inputs such as a keyboard, without retraining or fine-tuning the VLA. The method writes user input into the initial condition of the VLA's flow matching action expert; because the flow is deterministic, the generated actions both reflect user intent and stay close to the VLA's own unsteered actions.
DEFLECT is an offline post-training framework that turns the temporal gap between prediction and execution into preference supervision, enabling policies to produce actions aligned with the execution-time state even when their inputs are stale.
Researchers propose an interpretability pipeline to analyze what a fine-tuned vision-language-action (VLA) policy is computing internally. On pi0.5, injecting task A's activations into the action expert while the policy views task B's scene caused 99.6% of episodes (n=1968) to execute task A's motor trajectory, indicating the injected program is bound to absolute workspace coordinates and ignores the visible scene.
Why it matters: The study uses activation injection to break down the internal computation of VLA policies, offering a mechanistic explanation for understanding their perturbation brittleness.
The research team presents Kamino, a GPU-based physics solver implemented in Python using NVIDIA Warp and integrated into the Newton framework, designed for large-scale parallel simulation of heterogeneous robotic systems with strongly coupled kinematic and dynamic constraints, such as kinematic loops.
A study on deep reinforcement learning for legged robots proposes using latent actions learned from a small set of expert demonstrations as an inductive bias to help joint-level torque control policies explore more efficiently. This prior allows the agent to not be limited by the demonstration reward level and significantly improves performance on transfer tasks; when combined with style rewards for imitation, it can more closely match expert behavior. Related videos have been made public.
The paper proposes a model that can simultaneously co-generate video and 3D motion for any given skeleton, along with a camera that registers that motion to the scene. The model keeps 3D motion in a separate module to support arbitrary skeletons, connects it to a pretrained video model through a small number of cross-attention layers, and uses a generative video editor to curate nearly 200K training pairs. Evaluations show that its text-driven animation of diverse skeletons is more faithful than video-first pipelines, that the motion responds to the surrounding scene, and that it also supports video-to-skeleton motion, skeleton motion-to-video, and keyframe-based 3D motion control.
PUN (Planning Under Noise) is a neural operator that learns noise-aware value fields to enable motion planning with control execution noise in unseen geometric environments. It changes obstacle boundary conditions to reflecting Neumann conditions and sets a Dirichlet condition at the goal, and extends the Monte Carlo solver Walk on Stars (WoSt) to higher dimensions, learning the mapping from geometry to value field from noisy, partially converged labels.
CoGaussian proposes an open-vocabulary collaborative semantic occupancy prediction framework that uses sparse 3D Gaussians aligned with foundation models to represent scenes as a compact representation for transmitting and fusing geometric and semantic information among multiple agents, no longer relying on dense 3D voxel annotations and no longer limited to predefined semantic categories. On the Semantic-OPV2V benchmark, the method outperforms existing open-vocabulary single-agent approaches.
CLOSER is an evidence-aware active verification framework for object-goal navigation (ObjectNav) that treats stopping decisions as closed-loop evidence acquisition, combining vision-language model (VLM) verification with active perception.
A study proposes a distributed platoon controller that analyzes the asymptotic behavior of multi-vehicle traffic at an unsignalized intersection through closed-loop path modeling. It shows that under mild conditions on the controller gains, safe crossing is guaranteed during the transient phase, and in steady state all vehicles cross smoothly at a common constant speed while avoiding collisions. Unlike traditional bi-level decoupled designs on open paths, the method couples intersection control with platoon convergence analysis. Its effectiveness was validated in a high-fidelity CARLA simulator using multiple commercial vehicle models.
To address the increased collision risk when unmanned aircraft systems (UAS) resume autonomous navigation after temporary takeover by teleoperation or safety mechanisms, the study proposes three interruption-aware recurrent reinforcement learning architectures that incorporate interruption context into recurrent processing through independent feature streams, dedicated interruption memory, or gated environmental features.
The paper introduces RoboJudge-Bench to assess the reliability of multimodal large language models (MLLMs) as judges of embodied video generation, covering two dimensions, physical consistency and instruction alignment, with 800 human-annotated videos from 8 source corpora and 10 video generators, spanning grippers, human hands, and dexterous hands.
A study of 9 vision-language models, 682 images, and 7 evaluation conditions found that models' exact-counting ranges differ by nearly two orders of magnitude (R80 from 3 to 150 dots). Using a training-free agent loop that lets models mark images to aid counting, the study found this method mostly failed as an intervention but produced auditable serial reasoning traces that help reveal how models arrive at answers. Experiments also showed that models often change answers when traversal order is reversed, and that a child-like advantage from regular arrangements appeared only under specific conditions.
A controlled evaluation on the HighwayEnv testbed (1Hz, with gpt-4o-mini used for the key comparisons) shows that the published gains from experience memory in closed-loop LLM driving agents mainly come from sampling luck in the warmup pool: retrieval-based memory beat No-Memory in only one of ten pools (sign test p=0.011).
PanoGraphGS integrates panoramic 3D Gaussian representation learning and hierarchical 3D scene graph construction into a single framework, learning semantic and instance features on shared Gaussian geometry to achieve cross-view-consistent panoptic prediction and novel-view panoptic segmentation.
The paper proposes SynCalib, a data generation framework that addresses overfitting and poor generalization caused by insufficient extrinsic diversity in training data for deep learning-based online LiDAR-camera calibration. Using omnidirectional image datasets annotated with depth ground truth, it synthesizes multimodal calibration data under strict extrinsic constraints. Experiments show that this pipeline significantly improves the generalization and robustness of learning-based calibration networks, and real-world evaluations across various camera-LiDAR configurations validate the zero-shot transfer capability of models trained solely on synthetic data; the data generation code, pretrained models, and online benchmark dataset will be made public.
OmniVLoc proposes a surround-view visual localization system for omnidirectional cameras based on a topo-metric framework. It fuses global place recognition with monocular depth estimation to achieve multi-constrained localization, and solves a scale-aware pose graph and spherical bundle adjustment to recover metric-level poses. The system includes a hierarchical verification mechanism that filters erroneous outputs at each stage to improve robustness. Experiments on public datasets and a custom dataset validate its accuracy and reliability, and the source code and dataset will be released after the paper is published.
arXiv has released a monograph titled 'Physical AI: Concepts, Techniques, and Challenges,' comprising 12 chapters, 388 pages, and 64 figures. It defines physical AI as artificial systems that perceive, model, and act upon the physical world through continuous feedback under real-world constraints, and distinguishes it from digital AI, embodied AI, cyber-physical systems, and digital twins.
The paper proposes Latent Projection for Latent Planning, noting that planning costs in latent world models are often taken as the ℓ2 distance between the predicted terminal latent and the goal latent, while the latent itself contains noise whose source is precisely the regularizer used during training to prevent collapse.
The P3L (Preference-PCA Pluralistic Learning) framework jointly learns trajectory embeddings and a shared basis of preference directions via PCA decomposition from anonymous, heterogeneous crowd feedback, enabling personalized reward learning and downstream policy optimization without requiring annotator identities. In experiments across multiple embodied control tasks, P3L supports steerable personal alignment, generalizes beyond the offline dataset to out-of-distribution new users, and achieves competitive reinforcement learning and test performance while substantially reducing computational cost and supporting variable adaptation query budgets.
To address the limited surface area on thin cylindrical objects that makes planar markers difficult to place, a dual-color stripe marker is proposed. It uses the cylinder's known curvature to improve monocular camera pose measurement accuracy, with the two-color boundary providing a long centerline and dots and printed edges providing features at different depths.
AdaDex is a goal-conditioned reinforcement learning framework that turns a few seconds of human hand-object trajectory (from motion capture data or HOI reconstruction from in-the-wild RGB(D) video) into generalizable dexterous skills. Its core, Hybrid Dexterous Mimicking, trains a shared policy across two sets of massively parallel environments: environments with dense motion tracking objectives inject the human prior, while environments with noised initial states and goals and a sparse object-success reward force the skill to work even when it deviates from the reference trajectory.
The research team proposes UniWALM, a unified world-action-language model that learns video, action, and language simultaneously in a single shared representation, where each modality serves as both a conditioning signal and a supervision target.
Industrial ReMEmbR is a shared spatiotemporal memory system that extends NVIDIA ReMEmbR to robot fleets, reducing server-side computation by generating scene descriptions only on demand.
PhysCanvas is a state-consistent visuo-tactile world model for contact-rich robot manipulation that unifies multi-view RGB, tactile images, marker displacement, task state, and actions within a single vision space, sharing a video VAE and a pretrained generative backbone.
AgenticRoboRL proposes a framework for post-training robot planners across different robot harnesses. Its training recipe combines supervised fine-tuning with multi-turn reinforcement learning, and harness-specific adapters connect planner interaction records and task outcomes to a unified RL training pipeline.
SceneRig is an agentic system that turns a single image into an editable, simulation-ready scene. It combines geometric, visual, and physical tools to reconstruct 3D scenes and achieves physical stability through edit-and-validate loops inside a physics simulator.
An arXiv paper proposes MAPLE (Mondrian-Anchored Prior for Learning-free Indoor Exploration), a training-free predictive exploration method that adapts the Mondrian process into an online generative prior conditioned on geometric cues from the current partial map.
REACT is a framework for learning deployable humanoid-human interaction policies from human-human demonstrations, using semantic contact structure as the supervisory signal that links demonstration processing, interaction learning, and onboard control. It uses a VLM-guided contact modeling and retargeting pipeline to identify semantic contacts and transfer them onto humanoid-human geometry, then distills a privileged interaction policy into an egocentric depth policy that relies entirely on onboard sensing, and introduces a generative partner state modeling objective to predict partner motion from occluded viewpoints.
SuperMimic uses a differentiable world model of shared dynamics to train the world model and the tracking policy together in a single training loop, giving the policy analytic gradients through backpropagation-through-time over short imagined rollouts.
Why it matters: The paper lays out a concrete path for replacing large-scale reinforcement learning training with a single differentiable world model and provides compute comparison data, making it possible to assess its impact on scaling humanoid whole-body control.
The paper proposes FRDV (Fast Retargeting with Dynamics Validation), a framework for rapidly retargeting large-scale human motion capture data to different humanoid robots.
The authors present MARE, an open-source lunar rover perception and autonomy simulation suite built on Unreal Engine 5 and ROS 2, comprising three parts: terrain generation, dual-physics lunar rover sensor simulation, and GNC mission control. The simulator provides stereo, lidar, event, and SPAD camera data along with pixel-level ground truth; at low solar elevation angles, the mean IoU between predicted shadows and rendered shadows is 0.79. Stereo depth analysis shows that pixel coverage in shadowed regions is lower than in sunlit regions.
The research proposes GPT-as-Tactile-Policy, which summarizes expert demonstrations into task guidance and tactile references, enabling GPT-6-Astra to combine multimodal observations and interaction history to generate control commands and adjust actions based on contact feedback.
Idealworks and NVIDIA systematically evaluated the sim-to-real transfer of the X-Mobility and COMPASS learned navigation pipeline on industrial mobile robots. After training in a digital twin with Cosmos-generated sim-to-real data, real-robot success rates across four scenarios rose from 65.3% to 91.7%, weighted travel time nearly halved, the gap to the production-line navigation stack was 8.3 percentage points, and everything ran on the onboard Jetson AGX Xavier.
The paper proposes TaskWAM, a closed-loop, stage-aware framework that uses Mode Tokens to represent fine-grained task requirements and State Tokens to track execution progress, forming a structured task state that steers world-action model generation toward futures matching the current task, addressing the Stage Confusion phenomenon in long-horizon manipulation.
Researchers propose the CAPSafe benchmark, which evaluates the safety reasoning capabilities of large language model-driven embodied agents at two levels: an action-level dataset that judges whether a single action is safe in a given environmental context, and a plan-level dataset that assesses whether agents can generate safe long-horizon plans for complex hazardous household tasks.
The paper proposes SPAR (Search-trained Push Action Ranker), which learns to rank push primitives from simulated search results to alleviate the exponential growth of candidate action sequences with sequence length in Navigation Among Movable Obstacles (NAMO).
This study proposes Reinforced Iterative Learning Control (RILC), which couples Iterative Learning Control (ILC) with Reinforcement Learning (RL) for trajectory tracking of elastic-joint robots, addressing the problems that traditional ILC must learn each new trajectory separately and that RL has high sample complexity and limited tracking accuracy.
The paper systematically evaluates GPT-6 Astra's closed-loop robot manipulation capabilities using a lightweight agent execution framework, finding a 26.92% success rate on RoboDojo and 76.0% on RoboTwin 2.0, but a clear gap between semantic reasoning and physical execution.
The paper proposes a Safe Wrapper algorithm wrapped around robot communication that tests the stream of trust values output by each robot's anomaly detector using Safe Anytime-Valid Inference, thereby selecting trusted links whose data can be safely used for state updates.