UniWAM open-sources a unified world-action model hitting 99.2% on LIBERO with human-robot co-training scaling

机器之心 · wechat · 2026-10-13

Euler Wanxiang and HKUST(GZ) released UniWAM, the open-source version of Ola-Sem, unifying semantic understanding, future visual prediction, and action generation in one architecture — code and weights fully open.

Highlights:

Three components (Qwen3-VL-2B physical reasoner, Wan2.2-TI2V-5B world generator, action predictor) exchange information via joint multimodal attention.
A "physical language" supervision scheme describes action segments in natural language, converting egocentric human video into robot-usable learning signal; 10,000+ hours of robot and human data were cleaned and annotated.
Results: 99.2% avg success on LIBERO, 92.6% on LIBERO-Plus, 68.32% on RoboTwin 2.0 C2R — best among compared methods; real-robot success of 67.5% across four task types, with different-interaction success at 72.5% vs π₀.₅'s 40.0%.
History-conditioned flow matching: 1-step generation already hits 64.99% success; 4 steps within 0.06pt of 10 steps.
First Scaling trend evidence for human-robot co-training: log-linear loss vs. data scale (R²=0.983), with human data gains growing as robot data scales.
Positioned as groundwork for a Real-World RSI (recursive self-improvement) pipeline.

Original post →

UniWAM open-sources a unified world-action model hitting 99.2% on LIBERO with human-robot co-training scaling

More from Embodied