INTACT by ZJU & Tsinghua: Robots Skip Trial-and-Error to Act Directly on Intent

jiqizhixin · x · 2026-08-11

Researchers from Zhejiang University and Tsinghua University introduced INTACT, a novel method that trains world models to map directly from intent to action. This allows robots to skip expensive trial-and-error planning during test time.

Experiments show INTACT achieves a 98.7% success rate (outperforming LeWM CEM's 67%) across four robotics tasks, while running 300x faster at under 6 milliseconds per step.

Original post →

More from Embodied

Embodied channel →