Running vision-language-action (VLA) models asynchronously lets robots hide inference latency behind motion: the robot keeps executing the current action chunk while the next chunk is generated from an earlier observation. The trade-off is that by the time the new chunk runs, the robot or scene has shifted, leaving the action matched to a stale state.
Earlier fixes adjust training inputs or supervised targets based on expected delay, or patch actions during deployment, but they do not directly supervise how actions should change as the environment evolves. The DEFLECT authors argue that execution-time states, while unavailable during online prediction, are visible in recorded offline trajectories. They therefore propose an offline post-training scheme that turns this temporal contrast into preference supervision for a policy acting on stale inputs.
Across Kinetix, LIBERO, and three real-robot tasks, DEFLECT improved delay robustness over the evaluated VLA policies. At the longest supported delay, it increased success by 4.2 percentage points over VLASH on Kinetix and by 22.7 points over naive asynchronous execution with GR00T N1.7 on LIBERO. This summary is based solely on the arXiv abstract, which does not include full experimental details.