Saved in:
| Main Authors: | , , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2605.19294 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866909055629393920 |
|---|---|
| author | Zhu, Yixiang Chen, Yonghao Meng, Rui Guo, Jingyu Zou, Jiaxiang Yang, Zijie Wang, Taowen Chen, Xinyu |
| author_facet | Zhu, Yixiang Chen, Yonghao Meng, Rui Guo, Jingyu Zou, Jiaxiang Yang, Zijie Wang, Taowen Chen, Xinyu |
| contents | Vision-Language-Action (VLA) policies are typically deployed with asynchronous inference: the robot executes a previously predicted action chunk while the model computes the next one. This creates a prediction-execution misalignment: the chunk is conditioned on the observation taken before inference began, but executes in a physical state that has already drifted forward by several control steps; naive asynchronous rollover collapses from 89% to under 1% on Kinetix as the inference cycle covers up to seven control steps. We introduce DEFLECT, a fully offline post-training refinement that applies as a near drop-in upgrade to existing async-VLA stacks by converting latency itself into a label-free preference signal: counterfactual fresh/stale action pairs are constructed from a frozen reference policy and scored under the deployment-time conditioning via an implicit flow-matching likelihood-ratio surrogate, with no human labels, reward models, or online rollouts. DEFLECT substantially extends the usable delay envelope of async VLA control, with +6.4 success-rate gain in the high-latency regime (5-7 control steps), +4.6 when transferred to a real-scale VLA at the longest delay, and consistent improvements on two real-robot tasks (a bimanual conveyor pick-and-place and a reactive whack-a-mole). |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2605_19294 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | DEFLECT: Delay-Robust Execution via Flow-matching Likelihood-Estimated Counterfactual Tuning for VLA Policies Zhu, Yixiang Chen, Yonghao Meng, Rui Guo, Jingyu Zou, Jiaxiang Yang, Zijie Wang, Taowen Chen, Xinyu Robotics Artificial Intelligence Vision-Language-Action (VLA) policies are typically deployed with asynchronous inference: the robot executes a previously predicted action chunk while the model computes the next one. This creates a prediction-execution misalignment: the chunk is conditioned on the observation taken before inference began, but executes in a physical state that has already drifted forward by several control steps; naive asynchronous rollover collapses from 89% to under 1% on Kinetix as the inference cycle covers up to seven control steps. We introduce DEFLECT, a fully offline post-training refinement that applies as a near drop-in upgrade to existing async-VLA stacks by converting latency itself into a label-free preference signal: counterfactual fresh/stale action pairs are constructed from a frozen reference policy and scored under the deployment-time conditioning via an implicit flow-matching likelihood-ratio surrogate, with no human labels, reward models, or online rollouts. DEFLECT substantially extends the usable delay envelope of async VLA control, with +6.4 success-rate gain in the high-latency regime (5-7 control steps), +4.6 when transferred to a real-scale VLA at the longest delay, and consistent improvements on two real-robot tasks (a bimanual conveyor pick-and-place and a reactive whack-a-mole). |
| title | DEFLECT: Delay-Robust Execution via Flow-matching Likelihood-Estimated Counterfactual Tuning for VLA Policies |
| topic | Robotics Artificial Intelligence |
| url | https://arxiv.org/abs/2605.19294 |