PhysX-Omni models rigid, deformable, and articulated objects in one simulation-ready 3D generation framework, while PhysXVerse contains over 8.7K physical 3D assets across more than 2.9K categories.

Southeast University and Peking University researchers introduced MetaFine, a diagnostic meta-evaluation framework that tests fine-grained robot manipulation across understanding, perception, and behavior, and the article says traditional binary success metrics can overestimate fine-manipulation capability by up to 70%.

Xpeng presented its world-model stack at CVPR 2026, covering X-World, X-Foresight, and X-Cache; the article says X-Cache cuts about 70% of repeated computation, the second-generation VLA used over 4 trillion training tokens, and the in-car stack reduced inference latency to 80 ms.

NVIDIA Research presented three physical AI papers at CVPR: GraspGen-X was trained on 2 billion simulated grasps, LCDrive cuts reasoning tokens by about half versus text-based reasoning, and NitroGen trains embodied agents across more than 1,000 games and 40,000 hours of interaction.

Ant Lingbo and HKUST introduced LingBot-VA, an autoregressive video-action world model that unifies visual dynamics prediction and action inference, and the paper reports fine-tuning with 50 real-world demonstrations per task plus 92.0% and 91.1% success on RoboTwin 2.0 Easy and Hard settings.

Daxiao Robot and NTU introduced PhysX-Omni, a unified simulation-ready physical 3D generation framework for rigid, deformable, and articulated objects, with PhysXVerse covering 8.7K assets across 2.9K categories and PhysX-Bench evaluating six dimensions including geometry, scale, material, affordance, kinematics, and description.

Banafsheh Rafiee and Richard S. Sutton propose an enactive cognition framework for AI, naming four pillars: experience, perception-action inseparability, autonomy, and embodiment.

NVIDIA, Tsinghua, University of Toronto, and Vector Institute released Gamma-World, a multi-agent world model using simplex-based positional encoding and hub tokens to cut interaction cost from quadratic to linear, with 8-player latency dropping from 17.6 ms to 4.5 ms.

NVIDIA Research presented 8 ICRA papers on sim-to-real robotics: ScheduleStream delivered a 3x speedup for multi-arm planning, COMPASS reached about 80% success across 20 real-world navigation trials, and Grasp-MPC achieved about 75% real-robot grasping success.

NTU AutoMan Lab, Harvard, and Xiaomi Auto proposed AutoMoT, a unified VLA driving model using a 4B Qwen3-VL Understanding Expert and a 1.6B Action Expert with asynchronous inference, reaching 89.42 DS and 74.09% SR on Bench2Drive with AutoMoT+.

Samsung disclosed three AI efforts—Meki, M2RL, and LiveClawBench—covering a memory-based edge architecture, multi-domain reinforcement learning, and Physical AI evaluation; the article also says Samsung has purchased tens of thousands of GPUs for AI infrastructure, but does not provide model size, training budget, or deployment timelines.

Westlake University and collaborators introduced HiF-VLA, a motion-centric VLA framework that extracts compact Motion vectors with codecs such as H.264 and uses a joint expert to predict future visual motion and generate action sequences, reporting 31.4GB peak memory and 117.7ms latency under the cited history-window setting.

NTU MARS Lab and collaborators released a 43-page survey on robot world models, covering definitions, architectures, applications, benchmarks, and challenges around action-conditioned consistency, inference efficiency, and physical grounding.

PsiBot says it uses 100,000 hours of human operation data to train robot policies, with the W0 world model acting only as a training-time transfer module while deployment runs R2 alone.

Richard Sutton and coauthors proposed Intentional Updates, which derive the step size from the desired output change; Intentional AC approached SAC on MuJoCo under batch=1 streaming training without replay, while each update used about 1/140 of SAC’s FLOPs.

NVIDIA Cosmos Lab and Purdue University proposed Scenethesis, a language-and-vision agentic framework for text-to-3D scene generation that uses visual grounding, SDF-based physical constraints, and a judge module; experiments report about 72% first-pass success, 91% after self-checking, and collision rate reduction from 6.1% to 0.8%.

Jianlan Luo’s team and Agibot released LWD, tested on 16 Agibot G1 robots in real settings. LWD Online scored 0.95 across 8 tasks and 0.91 on long-horizon tasks. Its offline-to-online RL uses failures as data; failed trajectories were 34.8% of a 652.5-hour pool.

Over 10 universities released an 88-page survey proposing a “capability level × domain law” framework for world models. It reviews 400+ works and reports the best video models pass physical-consistency tests at only 26.2%. The key L3 case is A-Lab: 353 closed-loop experiments in 17 days, yielding 36 compounds.

NTU MARS Lab released REI-Bench, a benchmark with 9 ambiguity levels for vague human instructions. Tests used 4 robot planning frameworks and 6 small LLMs; LLaMA3.1-8B+SayCan fell from 57.7% to 46.9% in standard multi-turn context. The key issue is implicit reference resolution, where baseline success dropped 7.4% to 36.9%.

Psibot says it released a 100,889-hour human-plus-robot manipulation dataset, and that Psi-R2 ranked first on AllenAI’s MolmoSpace benchmark. The post lists 95,472 hours of human data, 5,417 hours of robot data, 1,000 open-sourced hours, 294 scenes, 4,821 tasks, and 1,382 objects; Psi-W0 adds 30% failure samples, and Psi-R2 latency drops from 2.2s to under 100ms. The key point is the data loop and benchmark framing: the post claims nearly 10x higher success, but does not disclose task setup, full baselines, or statistics.