AIBRIEF
MIT's robotics AI system VLASH plans a robot's next move while the current one runs, cutting reaction delays up to 11.8× on the same hardware. Demos are research benchmarks, not deployed robots.
A new system called VLASH, from a team at MIT, Nvidia, Caltech, UC Berkeley, UC San Diego, and Tsinghua, lets research robots finish tasks 1.5 to 2 times faster on the same hardware, with reaction delays cut up to 11.8 times. The fix is a scheduling trick, not a smarter model.
Today's AI-controlled robots stop and start. Each move waits for a vision-language-action (VLA) model, the kind of "brain" that turns camera input and language instructions into physical actions, to finish planning the next one. That stop-and-go is the bottleneck, and the wait can stretch into hundreds of milliseconds.
VLASH overlaps the next action's planning with the current action's execution. It predicts the robot's own future state, where the arm will be a few frames from now, so the next chunk of inference can start before the previous chunk finishes. The researchers say it avoids the heavier compute cost of building a full world model.
An earlier draft of the paper reported a roughly 30× speedup. The team told Live Science that figure used longer action sequences and slower hardware, and that ~12× on shorter, more realistic settings is the more representative number.
The work is a preprint on arXiv (2512.01031), with the code released on GitHub, and a presentation scheduled for the Intelligent Robots and Systems Conference this fall. Demos ran on controlled research benchmarks, not deployed robots.