MIT's VLASH Doubles Robot Speed by Predicting Its Own Next State

MIT researchers published VLASH, a technique that lets a vision-language-action model plan its next chunk of motion while the robot is still executing the current one. The trick is that the model does not need to know the future environment — it knows where the robot is and what it is about to do, so it can predict its own end state and start planning from there. That removes the stall between action chunks that makes VLA-driven robots visibly stutter, cutting reaction lag by more than 30× and roughly doubling speed on pick-and-place. Same policy, same hardware; the win is entirely in scheduling.