Researchers led by Umer Siddique have presented ReGround, a method that keeps vision-language-action models working when a robot's actuation changes after deployment, according to a paper accepted at a CoRL 2026 workshop and posted on arXiv. Remounting, mechanical wear or changed command latency can make such models fail tasks or behave unsafely, the authors write.
ReGround leaves the model unchanged. It learns the deployed robot's dynamics from random motor movements unrelated to any task, then tracks the motion the model intends while staying close to its plan. It switches on only when it detects a change, and needs no rewards, demonstrations or retraining.
Across two models, eight LIBERO tasks and five kinds of change, it brought success close to the ideal case; with 400 milliseconds of latency, for example, it raised the pi-0.5 model's success rate from 0.16 to 0.90, while staying safer than a linear calibration baseline.



