UNITAS: A 3D-Native World Action Model for Embodied Manipulation
UNITAS puts robot action and scene change into the same metric 3D frame.
The paper presents a 1.7B-parameter world action model for embodied manipulation, using 3D point trajectories for hands or grippers and scene-point displacement for the environment. It is built to handle both direct action execution and prediction of how an action changes a scene. The authors report up to 49% lower displacement errors than PointWorld on RoboTwin, plus 99.8% success on LIBERO and an 85% average across real-world tasks. Code is listed as available. HF Daily Papers' note
The paper presents a 1.7B-parameter world action model for embodied manipulation, using 3D point trajectories for hands or grippers and scene-point displacement for the environment. It is built to handle both direct action execution and prediction of how an action changes a scene. The authors report up to 49% lower displacement errors than PointWorld on RoboTwin, plus 99.8% success on LIBERO and an 85% average across real-world tasks. Code is listed as available. HF Daily Papers' note
score 6