Event

PACT-VLA: A Training-Free Action Chunk Verification Framework to Improve Asynchronous Inference Success Rate

1 reports · 1 sources

Yixiang Zhu et al. propose PACT-VLA, a training-free framework that, under asynchronous inference, scores candidate action chunks using pre-commit observations and selects the best-matching one, without requiring an independent verifier. Its scoring function, FlowRes, measures the residual between the direction of the candidate's implicit flow and the velocity predicted by the frozen policy under that observation; the pipeline implementation overlaps observation processing with candidate generation and performs scoring in batches, and on the evaluated hardware the selection does not add control-step latency. Reported results: under four-step latency, PACT-VLA achieves a 15.0 percentage point higher average success rate than RTC on Kinetix, and on LIBERO with GR00T N1.7 it achieves a 5.2 percentage point improvement over single-candidate asynchronous execution, and it further improves existing latency-robust methods; on three real-robot tasks, using the same frozen π0.5 checkpoint, it raises the average success rate from 60% for naive asynchronous execution to 82%.