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.
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.
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.
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 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.
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 authors propose Generative Plan Latents (GPL), which represent the reasoning state of vision-language-action (VLA) models as a low-dimensional conditional distribution over plans, such as which object, grasp, or route to use, so that plans can be sampled, selected on request, and reweighted by deployment outcomes without retraining the VLA.
The paper proposes a two-stage framework that learns complementary predictive representations from multi-view video clips to enhance embodied manipulation policies. In the pretraining stage, a frozen 3D foundation model provides dense geometric features, which are compressed into predictive geometric representations and optimized to preserve scene structure and its temporal evolution; action supervision is then applied to geometric transitions to extract a motor representation that emphasizes action-relevant transitions. In the downstream post-training stage, predictive geometry is transferred through geometry queries, and the motor representation guides intermediate features along the action-generation pathway. Future observations serve only as privileged supervision during training and are not needed at inference time.
The paper proposes DirectWAM, which converts pretrained video and action experts into single-step predictors that require no stochastic sampling while retaining the explicit future-to-action cascade structure. After learning a single-step action model, the method trains the future predictor by decoding actions from its own predicted futures and backpropagating demonstration action error.
PACT-VLA proposes a training-free framework that, in asynchronous inference, uses pre-commit observations to select the best-matching action chunk. Its scoring function, FlowRes, measures the residual between the direction of a candidate's latent flow and the velocity predicted by the frozen policy under that observation, without needing a separate verifier.
The study examines whether expected free energy (EFE) can determine how much computation to allocate to planning based on a learned world model, with an objective combining goal distance and disagreement among stochastic predictions. Experiments found that stopping when elite cost stalls reduces refinement while mean success rates stay close to fixed-depth references; with a frozen model, rewarding disagreement lowers success, while online updates partially recover losses caused by regions missing from the training data. The cost-stall rule fails when optimization exploits model errors, and predicted cost keeps falling after task success rates peak.
An arXiv paper presents LocoPilot, a single-operator bilateral teleoperation and data-collection framework deployed on a 12-DoF Unitree Go2 quadruped equipped with a 6-DoF AgileX Piper L manipulator, an Intel RealSense D435 wrist camera, and a D405 head camera.
The paper proposes the HuTact framework, which uses large-scale human hand-object interaction (HOI) data to pretrain tactile-aware hand representations. The pretraining stage requires no real tactile signals, and the model is then adapted with a small amount of unlabeled robot interaction data.
The paper proposes the PointHand teleoperation interface, which renders a multi-fingered robotic hand as sparse keypoints at the fingertips and joints, matching the keypoints observable by the policy. In a simulated mug-in-box task, 18 novices and 4 experts collected 2,200 trials across three rendering conditions; among them, novices who saw emphasized rendering of the thumb and index finger achieved 441 successful demonstrations out of 600 trials, higher than the 320 out of the real hand-shape rendering.
The paper conducts a cross-model evaluation of GPT-6 Astra, GPT-6.1 Sol, Claude Opus 5.5, and Gemini 4 Argon on RoboColiseum, covering four tracks: instruction following, spatial reasoning, general manipulation, and robustness.
The paper proposes a latent action model (LAM) that can ingest synchronized multi-camera views, using cross-attention to resolve depth and occlusion issues, then compresses them into latent actions that are fed into a policy in which a latent expert and an action expert are coupled by sequential flow matching around a frozen VLM backbone.
A CoRL 2026 paper proposes Advantage-Weighted Latent Steering (AWLS), an offline post-training method that improves frozen one-step generative manipulation policies without updating their decoders.
PLaCA proposes a world-to-action framework that directly predicts spatially structured patch-level future latent change from current multimodal input and explicitly incorporates it into action generation. On LIBERO-Long, PLaCA achieves a 79.0% success rate, a 5.0 percentage point improvement over the Action-only baseline; on CALVIN, the average sequence length improves from 2.784 to 3.564. The second-stage training updates only 2.95% of the model parameters.
A CoRL 2026 paper proposes the Action Influence Maximization (AIM) algorithm, which actively collects training data to increase the action influence of the dataset, thereby improving the controllability of action-conditioned video world models. The authors prove that a data's action influence metric upper-bounds the action influence any model trained on that data can express, and that the narrow action distribution of expert demonstrations causes this ceiling to be low.
The paper defines the problem of failure attribution for vision-language-action (VLA) models: when multiple visual corruptions coexist, determining which corruptions contribute to failure risk and how much benefit their removal would bring. The authors propose a factorized risk model that learns separate representations of the scene and of corruptions, and estimates removal benefits through latent intervention using only the current image and task instruction, without paired clean images or additional counterfactual rollouts.
EmbodyHOI runs a differentiable inverse kinematics (IK) solver for a specific humanoid embodiment inside the sampling loop of a frozen flow-matching hand-object interaction (HOI) trajectory generator, placing embodiment reachability constraints within the generation process rather than in post-hoc retargeting.
A study proposes a plug-and-play auxiliary loss called score-decorrelation to address the problem that diffusion policies with linearly composed scores tend to have convergent component predictions that are hard to interpret in multitask robot learning. The loss encourages specialization by penalizing overlap between component noise predictions, with almost no architectural changes and very low computational overhead. Experiments on multiple benchmarks show that the method consistently reduces duplicate counting between components, improves task decomposition, and produces behaviorally meaningful specialization without compromising multitask performance.
The authors propose Embodied Playground, a benchmark for evaluating how foundation-model agents coordinate physical, strategic, and spatial reasoning across long-horizon tasks. It contains 410 task environments spanning basic manipulation, sequential puzzles, and multi-stage household activities.
DexGraspFlow proposes a conditional link-level SE(3) flow-matching framework for cross-embodiment dexterous grasp generation. It combines geometric relations with kinematic topology, uses a dual-stream DiT to model link-link coordination and link-object interaction, and provides explicit structural information through attention biases derived from kinematic-tree distances.
The article analyzes why the dexterous hand industry is collectively reducing degrees of freedom: manufacturers are now talking about lifespan, cycle time, and unit price rather than making hands more human-like. A new hand from a leading US robotics company has only four fingers and 13 degrees of freedom, while two new domestic products have 22 and 20 degrees of freedom, compared with benchmark products from just over a year ago at 38, 37, and 42.
The paper proposes Video-Point Model (VPM), a world model that conditions on multi-view video histories and language instructions to jointly predict future videos from arbitrary viewpoints and the 3D trajectories of persistent scene points. The authors add camera-conditioned multi-view processing and a trajectory prediction head to a pretrained world model, and build the VPM-Data simulation corpus and the VP-WorldBench benchmark; they further add robot actions as a token modality in the generator, so that the policy shares the point-supervised backbone.
Under the Gravity framework, Geek+ embodied intelligence proposes the Gravity Control causal intervention model. Without training the original policy model or adding new demonstration data, it uses mechanistic interpretability to uncover the model's internal action regulation mechanisms and intervenes on relevant latent features at inference time to achieve runtime correction.
On October 14, RoboParty released the developer-oriented humanoid robot RP1 on X. As of October 20, the release post had about 49,000 views and 52 replies, and the official video had over 1.6 million plays. RP1 fully open-sources materials including mechanical CAD, BOM, the self-developed ROMOMO joints, firmware, motion control, simulation models, and SDK. It stands 1.35 meters tall, with 24 active degrees of freedom in the standard version and 30 in the Pro version; some joints have a peak torque of 160 N·m, and its top running speed is 4 m/s.
At its 2026 Global Innovation Conference, Chery announced the "Yaoguang 2035" strategy, declaring its transition from a traditional automaker to an AI-driven global intelligent technology company, and released CheryMind, a vertical-domain foundation model for physical AI.
Genrobot.AI reposted and congratulated Dream Labs on launching Dream-1, saying it is a launch partner for Dream-1. The quoted content says Dream Labs has exited stealth and released Dream-1, a robot foundation model that can control dual-arm robots it has never seen before, complete tasks that were not demonstrated, and work in scenes it has not visited, at a price of 3 USD per hour, with the Dream API opening to everyone that day.
CausalAlign proposes a causal alignment framework that uses linear probes and causal interventions to locate the representation layers for camera ego-motion in video generation models, then aligns the representations at those layers to a 3D foundation model via delta-token space. After fine-tuning Wan2.2 on DL3DV and RealEstate10K, geometric consistency improves by more than 20% over the base model, and on RoboCasa365 mobile manipulation tasks it raises the success rate of the world action model (WAM) by 10 percentage points.
The paper proposes LaLa-World, a language-conditioned latent world model (LWM) that learns jointly from simulation and robot trajectories as well as broader image and video data. Its key design is to describe both numerical actions and visual changes in language, enabling training of a single future predictor across heterogeneous datasets, and to build the predictor on a language model so it can interpret language conditions.
CausGround proposes a non-parametric causal interaction grounding framework that distills interaction experience into action-centered precondition and effect knowledge, and combines it with current observations during planning for backward and forward reasoning.
TARSphere proposes a full-stack framework for robots that evaluates long-horizon embodied agents under physical and social constraints. Its TARSphere-Infra runs robot policies in a physics simulator with embodied NPCs and constrained dialogue, and uses an event-driven finite-state machine to update task states based on completed actions and delivered information; TARSphere-Harness lets a frozen multimodal large language model revise its decisions through multi-level executable skills, observations, hierarchical memory, and execution feedback.
The authors deployed decentralized multi-agent reinforcement learning (MARL) on two physical robots for joint task assignment and navigation in a setting where multiple robots serve multiple production machines. Policies were first trained in high-fidelity simulation in Isaac Lab/Isaac Sim, then transferred zero-shot to the real robots via a self-built ROS 2 deployment pipeline. Paper: https://arxiv.org/abs/2609.14567.