Event

VLA policies still initiate execution without instructions; startup suppression methods are limited

1 reports · 1 sources

Minsu Jang and other authors released a study examining whether vision-language-action (VLA) policies mistakenly initiate execution when instructions are missing or meaningless. The study applied five types of instructions—empty, neutral, meaningless, infeasible, and wrong task—to six model checkpoints in the LIBERO-Goal simulation, fixing the initial state and changing only the instruction, to examine whether the robot starts, which task it completes, and what the instruction determines. The authors state that the fine-tuned policies never completed the task under non-task inputs, but could not reliably keep the robot stationary, nor could they suppress initiation under infeasible instructions. Among the three startup suppression methods, the prefix method failed to stop the robot in all six checkpoints; the gating method maintained suppression for non-task inputs in five of the six checkpoints.