Learning Native Continuation for Action Chunking Flow Policies

In three linesLegato is a training-time continuation method for action-chunked flow-based VLA policies. It initializes denoising from a mixture of known actions and noise, and reshapes flow dynamics to ensure consistency between training and inference. Real-world experiments: ~10% improvements in trajectory smoothness and task completion time versus RTC across five manipulation tasks.

Summary generated by Claude — human-verified