DexForce's UNITAS: a 1.7B 3D-native world action model tops LIBERO at 99.8%, beats 8B rivals
机器之心 · wechat · 2026-10-12
DexForce (跨维智能) unveiled UNITAS, billed as the first task-native 3D world action model for robotics. Instead of translating between 2D vision and action spaces, it unifies observation, action, and scene dynamics in one 3D physical representation: actions are point trajectories on the robot (Action Flow), environmental change is scene-point trajectories (Scene Flow).
Key results at only 1.7B parameters:
- 99.8% average success on LIBERO, 89.1% on LIBERO-Plus
- 92.94%/93.04% on RoboTwin 2.0 dual-arm (Clean/Randomized), beating 6B Fast-WAM, 8B Motus, π0.5, and others
- Scene-motion prediction error cut from 0.656cm to 0.427cm; 85% average success on real-robot tasks
Ablations show removing the 3D-native designs drops LIBERO-Plus success from 86.47% to 67.58%. Mixed pretraining across nine sources (sim, real robots, human hands — all as point flows in one coordinate system) adds 2.6 points on LIBERO-Plus.
A companion module, TaskAnchor, grounds task state from execution history: plugged into π0.5 it lifts RMBench success from 14% to 79% with only 27M extra parameters and 2ms latency. Code and papers are open-sourced.
More from Embodied
- Reflex teaches humanoid robots to catch thrown boxes from egocentric vision, 65% success on real Unitree G1 — UW · 2026-10-12
- RealHorse app quantifies horse lameness from smartphone video, dubbed 'Horse Tinder' — RexDouglass · 2026-10-12
- Yeah Hand: open-source humanoid robotic hand kits start at just $217 — CyberRobooo · 2026-10-12
- Fly-by-Code: coding agents program a tiltrotor drone with an arm, sim-to-real — furongh · 2026-10-12
- StreamRender-H3 renders interactive game worlds in real time with a 2-step diffusion model at 24 FPS — songhan_mit · 2026-10-12
- Artisan: multi-agent pipeline generates physically grounded articulated 3D assets — nanyang-technological-university-singapore · 2026-10-12