CurrentWorld-0: A robot world simulator combining cross-embodiment, multi-view and force-tactile prediction

机器之心 · wechat · 2026-08-17

CurrentRobotics has released CurrentWorld-0, an interactive world simulator that learns physics from real interaction data instead of hand-written physical formulas, and can be used for robot simulation, evaluation, and training.

Its capabilities center on three directions: cross-embodiment (fixed-arm, mobile, and humanoid robots keep their own action subspaces while sharing the same world physics), multi-view (head/wrist/third-person cameras are jointly modeled and anchored to one physical state), and force-tactile prediction (predicting contact forces alongside video, so grasp stability and task success can be judged).

The team notes that strong video priors can silently 'fix' robot mistakes—rendering a successful outcome when the grasp actually failed—so action controllability is essential (extending their ICML 2026 Spotlight dWorld Eval). Building on Hi-WM (RSS 2026 workshop), failures become explorable data nodes: humans can teleoperate corrections, and intermediate states can be saved, rolled back, and retried, closing the loop of real experience → learning → world-model evaluation → human correction → retraining. Demos include folding socks, spreading bread, peeling cucumbers, and pouring beer with deformables and fluids.

Original post →

More from Embodied

Embodied channel →