Event

μ0: Replacing Pixel and Action Modeling with 3D Interaction Trajectory Prediction as a World Model

1 reports · 1 sources

Seungjae Lee et al. proposed and experimentally validated the scalable world model μ0 and the accompanying TraceExtract supervised extraction system; the paper has been accepted to CoRL 2026. μ0 does not predict dense pixels, nor does it directly model actions; instead, it predicts smooth 3D trajectories of salient interaction points such as objects, tools, hands, and contact regions, forming a compact, embodiment-agnostic motion interface. Because μ0 can be frozen and reused, it can be paired with action experts and adapted to different robot embodiments. The authors state that, although pretraining does not use action data, the performance of its trajectory-conditioned policy remains comparable to VLA models with action-supervised pretraining, such as π0.