Vision-language-action policies often run asynchronously: the next action chunk is generated while the robot is still executing the previous one. By the time the new chunk is committed, the observation it was conditioned on is stale, and task success drops as inference delay grows. Existing multi-candidate approaches either rank candidates using the same stale observation or switch between candidates during execution, risking discontinuous action sequences.
PACT-VLA treats the inference delay as decision time. It generates candidate chunks from the inference-start observation and, while generation is still underway, uses the VLA itself to pick the complete chunk that best fits a later pre-commit observation. The scoring function, FlowRes, computes the residual between each candidate's implied flow direction and the velocity predicted by the frozen policy under that newer observation, so no separate verifier is required. A pipelined implementation overlaps observation processing with candidate generation and batches scoring, adding no control-step delay on the evaluated hardware.
At a four-step delay, the authors report that PACT-VLA improves average success by 15.0 percentage points over RTC on Kinetix and by 5.2 points over single-candidate asynchronous execution with GR00T N1.7 on LIBERO, while also improving existing delay-robust methods. On three real-robot tasks, it raised average success from 60% to 82% over naive asynchronous execution using the same frozen checkpoint. This digest is based on a single arXiv preprint, so there are no independent results to compare.