Reliability and safety
What a Last-Moment Check Reveals About Delayed VLA Control
Controlled-delay comparisons show whether a newer observation can prevent stale VLA action chunks from reducing task success, while keeping the policy frozen.
With asynchronous inference, the robot continues executing one chunk while calculating the next. The robot or its surroundings may change before that calculation finishes, leaving the next chunk tied to an observation that is no longer current.
What the test measures
At a fixed asynchronous delay, a pre-commit test can reveal reliability by measuring whether selection against a newer observation produces a higher task-success rate than committing a chunk based only on the stale inference-start observation. In PACT-VLA, multiple complete candidate chunks are generated from the earlier observation, but the final choice is deferred until a later observation is available just before commitment.
FlowRes compares each candidate's implied flow direction with the velocity predicted by the same frozen policy from the pre-commit observation. This requires neither policy retraining nor a separate verifier. Selecting a complete chunk also avoids the repeated switching between candidates that the authors identify as a potential source of discontinuous action sequences.
The authors report that their pipelined implementation overlaps observation processing with candidate generation and batches the scoring work. They say this selection added no control-step delay on the hardware they evaluated, a result that should not be generalized to untested hardware.
What the reported comparisons show
At a four-step delay, the authors report that PACT-VLA increased average success by 15.0 percentage points over RTC on Kinetix. On LIBERO with GR00T N1.7, it improved average success by 5.2 points over single-candidate asynchronous execution.
Across three real-robot tasks, the preprint reports average success rising from 60% under naive asynchronous execution to 82% with the proposed method using the same frozen π0.5 checkpoint. Holding the checkpoint constant makes the comparison more directly about execution-time selection, but the available abstract does not supply task-level results, trial counts or uncertainty estimates.
These tests treat reliability as retaining task success when inference introduces a known lag. The reported results support the value of a newer pre-commit observation within Kinetix, LIBERO, three physical tasks and the evaluated hardware, but they are claims from one preprint rather than independent confirmation across VLA policies and robot platforms.