Action-conditioned robotics world model fine-tuned from Cosmos-Predict2.5-2B using 2D skeleton conditioning across embodiments.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
OSCAR is an omni-embodiment, action-conditioned world model for robotics. It fine-tunes Cosmos-Predict2.5-2B with a unified 2D kinematic skeleton representation as the control signal, covering four robot embodiments and a MANO human hand. Training uses curated robot teleoperation data together with egocentric human video.
The model predicts the visual outcome of actions given a skeleton-style motion input, so the same action representation can transfer across different robot bodies. The project page lists sections on policy evaluation, a data processing pipeline, comparisons with baselines and ablations. A creator review notes the code is released under Apache 2.0 and that running it needs roughly 24GB or more of VRAM.
The work is by Zhuoyuan Wu and Jun Gao (arXiv 2606.04463, 2026). It is a research release aimed at robotics researchers, for example for evaluating policies in a learned simulator. The vendor or affiliation is not stated in the provided material, and the code and weight links were not found on the fetched page.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
World model for robots that predicts outcomes of actions using 2D skeleton-style motion as control, transferring across robot bodies. Code released under Apache 2, needs 24GB+ VRAM.
Closest neighbours by meaning, not just by category.
Model
Research framework that conditions one video world model on force, camera, hand pose and robot pose actions.
Model · Dyna Robotics
World-action model from Dyna Robotics pre-trained on over 1 million hours of egocentric human video for robot manipulation.
Model · Black Forest Labs
7B open-weights world action model from Black Forest Labs that jointly predicts video and actions for robot control.
Library
Embodied harness from NUS Show Lab that lets vision-language models control robots via discrete semantic action units.
Research · BitRobot Foundation
Open dataset of 500+ hours of whole-body teleoperated Unitree G1 humanoid data recorded in real homes.
Model · NVIDIA
NVIDIA's open omni-model family for physical AI: world simulation, reasoning and action generation for robots and AVs.
