FocalPolicy: Frequency-Optimized Chunking and Locally Anchored Flow Matching for Coherent Visuomotor Policy

Researchers introduced FocalPolicy, a new visuomotor policy framework for robotic manipulation.

It adds a foresight-oriented objective over action chunks and uses locally anchored sampling to train more efficiently.

The method improves cross-chunk consistency, helping produce smoother, more coherent long-horizon actions.

Experiments show FocalPolicy outperforms prior approaches on imitation learning and control tasks.