Researchers have developed an AI system that enables robots to plan their next actions while still completing their current movements. Called VLASH, or Vision-Language-Action with Scheduled Heuristics, the technique addresses a major limitation of robots controlled by vision-language-action models: the pauses that occur while the AI calculates what to do next, tells Live Science.
Conventional VLA-controlled robots often follow a stop-and-go pattern. After executing one set of instructions, they wait for the model to generate another. VLASH removes much of this delay by estimating where a robot will end up after its scheduled movements and planning the next actions from that predicted state. Tests showed that robots completed some tasks 1.5 to 2 times faster while maintaining most or all of their accuracy. Maximum reaction latency decreased by as much as 11.8 times, depending on the hardware.
Unlike world models that attempt to predict changes across an entire environment, VLASH predicts only the robot’s future state. Because it uses movements already scheduled, the system does not require a separate prediction model or additional inference step during operation. Researchers also reorganized training data, making individual fine-tuning steps 3.26 times faster in one benchmark.
The team evaluated VLASH on two robot platforms using tasks including picking and placing objects, stacking, sorting, table tennis, and Whac-a-Mole. The system processes new observations 15–30 times per second, helping robots respond when conditions change.
Faster reactions, however, do not automatically make robots safer or more intelligent. Researchers still need to evaluate VLASH around people, across more robots and environments, and under unexpected disturbances. Its first practical benefits could emerge in manufacturing and other applications requiring continuous motion and rapid responses.
