DEFLECT: Temporal Counterfactual Preference Learning for Delay-Robust Asynchronous VLAs

Name: DEFLECT: Temporal Counterfactual Preference Learning for Delay-Robust Asynchronous VLAs
Uploaded: 2026-10-22T16:58:10.242468+00:00
Duration: 819 s
Description: DEFLECT teaches asynchronous Vision-Language-Action (VLA) policies to prefer actions aligned with execution time over stale observations. It achieves this by creating a 'counterfactual preference' where chunks from future execution time are preferred over those from prediction time. This extends the usable delay envelope for VLAs, improving performance significantly in high-latency scenarios.

DEFLECT teaches asynchronous Vision-Language-Action (VLA) policies to prefer actions aligned with execution time over stale observations. It achieves this by creating a 'counterfactual preference' where chunks from future execution time are preferred over those from prediction time. This extends the usable delay envelope for VLAs, improving performance significantly in high-latency scenarios.

Open the full episode