Unitas: A 3D-Native World Action Model for Embodied Manipulation

Ruixiang Wang^1,2*,   Yongyi Su^3*†,   Wenlve Zhou³,   Bo Yue^1,2,   Hengyan Liu^1,2,   Dekun Lu²
Yuxin Tian^1,2,   Yihan Fang^1,2,   Zerui Wu²,   Xing Hu²,   Jietao Chen²,   Yong Guo²
Ziyan He²,   Junbin Yuan⁴,   Guiliang Liu¹,   Xiaofen Xing⁵,   Kui Jia^1,2

_{¹ CUHK-SZ   ² DexForce   ³ FSU   ⁴ SYSU   ⁵ SCUT
^* Equal contribution   ^† Correspondence to <suyongyi@fosu.edu.cn>}

Overview

UNITAS is a 3D-native world action model for embodied manipulation. It unifies observations, actions, and scene dynamics in a shared metric 3D frame within each interaction, using a common representation across robot embodiments and human hands.

Highlights

• One physical space. Observations, actions, and scene dynamics share a metric 3D world-coordinate frame.
• A common motion representation. Point trajectories represent robot motion, human hand motion, and scene dynamics.
• Complementary capabilities. The framework supports robot-action generation and action-conditioned scene prediction.
• Strong results with 1.7B parameters. Success rates reach 99.8% on LIBERO, 89.1% on LIBERO-Plus, 56.2% on VLABench, and 92.94% / 93.04% on RoboTwin Clean / Randomized. Moving ADE is 35% lower than PointWorld on RoboTwin scene prediction.

Release Plan

We are preparing the following releases:

• Model checkpoints and inference code
• Training code
• Training dataset