Shanghai AI Laboratory introduced InternW0, the first model in its InternW physical world model series, published as an arXiv paper. It jointly learns future visual dynamics and continuous robot control via an asymmetric video–action architecture with flow matching, trained on ~7,200 hours of heterogeneous robot and egocentric data. The work targets efficient, actionable world modeling for real-world embodied and scientific tasks.
Researchers present Agile-WAM, a tactile World Action Model that encodes visual and tactile observations into a shared latent driving a direct vision-tactile-to-action flow-matching process, jointly generating action chunks and future visual/tactile latents. Across nine simulated and five real-world contact-rich manipulation tasks it outperforms the strongest baseline in success rate while keeping inference latency at 11.9 ms, showing multimodal WAMs can run without large pretrained generative backbones.
Researchers present Agile-WAM, a tactile World Action Model that encodes visual and tactile observations into a shared latent driving a direct vision-tactile-to-action flow-matching process, jointly generating action chunks and future visual/tactile latents. Across nine simulated and five real-world contact-rich manipulation tasks it outperforms the strongest baseline in success rate while keeping inference latency at 11.9 ms, showing multimodal WAMs can run without large pretrained generative backbones.
A new arXiv paper introduces DexTacWAM, a World-Action Model that injects fingertip tactile latents into a video diffusion world model for joint visuo-tactile prediction. Across six contact-rich dexterous manipulation tasks on a 22-DoF bimanual robot, it tops every task, averaging 70.6 versus 38.0 for the strongest baseline. The result matters to robotics practitioners because it shows pretrained video priors can extend to touch with roughly 100 demonstrations per task and no tactile midtraining.
A new arXiv paper introduces DexTacWAM, a World-Action Model that injects fingertip tactile latents into a video diffusion world model for joint visuo-tactile prediction. Across six contact-rich dexterous manipulation tasks on a 22-DoF bimanual robot, it tops every task, averaging 70.6 versus 38.0 for the strongest baseline. The result matters to robotics practitioners because it shows pretrained video priors can extend to touch with roughly 100 demonstrations per task and no tactile midtraining.
An arXiv survey introduces a capability hierarchy for world models in embodied intelligence: Plausible, Controllable, and Actionable. It argues evaluation should shift from visual fidelity to whether predictions improve closed-loop behavior, and surveys manipulation, navigation, locomotion, autonomous driving, and general embodied learning.
An arXiv survey introduces a capability hierarchy for world models in embodied intelligence: Plausible, Controllable, and Actionable. It argues evaluation should shift from visual fidelity to whether predictions improve closed-loop behavior, and surveys manipulation, navigation, locomotion, autonomous driving, and general embodied learning.
Researchers present ABot-World-0, an action-conditioned video world model capable of real-time, long-horizon closed-loop interaction on a single desktop GPU. The model is trained on a multi-source data pipeline spanning AAA games, simulation engines, and internet videos, using agent-driven data collection and a unified quality assurance process.
Researchers present ABot-World-0, an action-conditioned video world model capable of real-time, long-horizon closed-loop interaction on a single desktop GPU. The model is trained on a multi-source data pipeline spanning AAA games, simulation engines, and internet videos, using agent-driven data collection and a unified quality assurance process.
A new arXiv paper introduces CLAP, a framework for cross-embodiment action-conditioned video generation trained on diverse internet-scale videos spanning human and robotic agents. It targets the limitation of existing action-conditioned video models that are restricted to a single robot embodiment. The work matters to robotics practitioners because it aims to turn heterogeneous video data into generalizable, zero-shot physical simulation.
A new arXiv paper introduces CLAP, a framework for cross-embodiment action-conditioned video generation trained on diverse internet-scale videos spanning human and robotic agents. It targets the limitation of existing action-conditioned video models that are restricted to a single robot embodiment. The work matters to robotics practitioners because it aims to turn heterogeneous video data into generalizable, zero-shot physical simulation.