Right Goal, Wrong Body: Re-grounding Frozen VLAs under Actuation Shift
Organizations: Department of Electrical and Computer Engineering University of Texas at San Antonio, United States
Abstract
Vision-language-action (VLA) models are capable of producing desired robotic behaviors in ideal settings. However, they often fail when the robot's actuation changes at deployment. For example, remounting, mechanical wear, or command latency changes can lead to failed tasks and unsafe behaviors. To address the inevitable actuation shift, we propose ReGround, an inference-time method that keeps the VLA frozen and re-grounds its intended motion. ReGround learns the deployed robot's dynamics from task-agnostic motor babbling, and then tracks the policy's intended motion while staying close to the policy's own plan. ReGround activates only upon detecting a shift, and needs no rewards, demonstrations, or retraining. We show via two VLAs, eight LIBERO tasks, and five shift settings that ReGround can restore near-oracle success, such as raising 's success from 0.16 to 0.90 under 400 ms latency, while remaining safer than a linear-calibration baseline. Project website and open-source code: https://umersiddique94.github.io/ReGround/.
Figures & tables
| Success | Unsafe-episode rate | ||||||||
| Shift | Base | Linear | Intent | ReGround | Oracle | Base | Linear | Intent | ReGround |
| Latency 250 ms | 0.61 | 0.49 | 0.80 | 0.96 | 0.97 | 0.68 | 1.00 | 0.17 | 0.29 |
| Latency 400 ms | 0.16 | 0.07 | 0.16 | 0.90 | 0.91 | 0.86 | 1.00 | 0.45 | 0.35 |
| Composite | 0.00 | 0.86 | 0.67 | 0.97 | 0.92 | 0.78 | 0.23 | 0.15 | 0.17 |
| Rotation 90 ∘ | 0.00 | 0.98 | 0.98 | 0.98 | 0.98 | 0.80 | 0.76 | 0.36 | 0.29 |
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
| latency 5 | latency 8 | composite | frame rot. 90 ∘ | gain 3 | ||||||
|---|---|---|---|---|---|---|---|---|---|---|
| task | base | RG | base | RG | base | RG | base | RG | base | RG |
| spatial_0 | 0.23 | 0.99 | 0.00 | 0.91 | 0.03 | 1.00 | 0.00 | 0.99 | 0.09 | 1.00 |
| spatial_2 | 0.54 | 0.95 | 0.00 | 0.83 | 0.00 | 0.98 | 0.00 | 1.00 | 0.03 | 1.00 |
| object_1 | 0.51 | 0.99 | 0.14 | 0.89 | 0.00 | 0.99 | 0.00 | 1.00 | 0.14 | 0.97 |
| object_3 | 0.66 | 0.95 | 0.11 | 0.87 | 0.00 | 0.96 | 0.00 | 0.99 | 0.03 | 0.95 |
| goal_1 | 0.74 | 0.95 | 0.40 | 0.93 | 0.37 | 0.95 | 0.00 | 0.99 | 0.34 | 1.00 |
Test-Time Adaptation of VLA Models
Test-time adaptation of vision-language-action models helps robot policies adjust to new views, tasks, objects, and execution errors during deployment.
See all 60 papers