UNITAS: a 1.7B 3D-native world action model hits 99.8% on LIBERO and 85% on real tasks
CUHKSZ · hf · 2026-10-14
CUHKSZ and DexForce introduce UNITAS, claimed to be the first 3D-native world action model. Unlike prior WAMs built on video generators, it unifies observations, actions, and scene dynamics in a shared metric 3D frame across robot embodiments and human hands: action flow represents grippers/hands as 3D point trajectories, scene flow models conditioned scene-point displacements, with world-aligned 3D positional embeddings (depth optional) and a physical-time trajectory tokenizer encoding each trajectory as one token. The 1.7B model cuts displacement error by up to 49% vs PointWorld on RoboTwin, achieves 99.8% on LIBERO and 85% average success on real-world tasks. Code open-sourced at github.com/DexForce/UNITAS.
Related event: CUHKSZ and DexForce Release UNITAS, a 3D-Native World Action Model(2 posts)→
More from Embodied
- EU privacy watchdog calls for smart glasses to clearly signal when recording — Polymarket · 2026-10-14
- How Argon drives Spot: real2sim skill verification in MuJoCo and map-free navigation — ki_ki_ki1 · 2026-10-14
- Agent Argon reads Spot's SDK docs itself, builds trash-sorting skill on Boston Dynamics robot — chris_j_paxton · 2026-10-14
- Meta researcher recruiting multimodal and robotics interns for next summer, no background needed — liuzhuang1234 · 2026-10-14
- Frontier LLM scores on real robot benchmark double every 6 weeks, 15% to 48% in 3 months — ycombinator · 2026-10-14
- Raven Prism AI glasses launch set for October 28 in San Francisco — smtabatabaie · 2026-10-14