A robot swaps its body and the frozen VLA still works: ReGround lifts success from 0.16 to 0.90 at 400 ms latency
Synopsis
When a vision-language-action (VLA) model is deployed on a robot whose actuation has changed, through remounting, mechanical wear or altered command latency, the same action command produces a different physical motion and tasks fail. ReGround keeps the VLA frozen at inference time, reads out the trajectory the policy intends using a forward model of the training body, then tracks that trajectory with a deployed-body model learned from task-agnostic motor babbling, while a trust region keeps actions close to the policy's own plan. Across two VLAs, eight LIBERO tasks and five shift settings it recovers 0.91–1.00 of lost success, for example raising success from 0.16 to 0.90 under 400 ms latency.
Interpretation
A frozen VLA can still complete tasks after the robot's actuation changes, provided its actions are re-grounded onto the deployed body at inference time without updating model parameters. Previously, handling this actuation shift generally required retraining or fine-tuning the policy, which needs additional task demonstrations, computation and access to model weights; existing test-time methods rely on reward, value or verifier signals to guide action selection and were not designed for frozen VLAs that output action chunks under actuation shift. Across eight LIBERO tasks and five shift settings, ReGround recovers 0.91–1.00 of lost success, reaching overall success of 0.90–0.98; it raises success from 0.16 to 0.90 under 400 ms latency, beats the best baseline by 0.16 at 250 ms latency and by 0.11 on the composite shift.
The method separates what motion the policy wants from how this robot executes it: the former is read out by a forward model of the training body, the latter is learned from data collected on the deployed body. The training-body forward model is fitted once on nominal babbling, while the deployed-body model comes from smooth Ornstein–Uhlenbeck motor babbling, optionally mixed with VLA-driven transitions, and needs no task labels, rewards or demonstrations. At 250 ms latency, 1000 babbling transitions (about 50 s of robot time) already give 0.98 success, where the one-step inverse-dynamics transfer needs 10000; the whole procedure uses no gradients, rewards or demonstrations.
Re-grounding switches on only when an actuation shift is detected, leaving the policy's actions unchanged on an unshifted body. Previously there was no mechanism ensuring that a test-time intervention does no harm when nothing has changed. A one-sided CUSUM test accumulates the normalized one-step residual of the training-body model, with a threshold calibrated on nominal validation episodes; on an unshifted body, always-on and monitor-switched settings give 0.98 versus 0.97 success at the same unsafe rate of 0.40–0.41, and the monitor catches every shifted episode after the shift starts while false-alarming on 2.1% of unshifted ones.
Perspective
The demonstrated scope is two VLAs, eight LIBERO tasks and five shift settings (action-frame rotation, controller gain, latency and a composite shift), controlled at 20 Hz in MuJoCo with robosuite, with hyperparameters selected on validation states 0–14 and reported on test states 15–49.
On static re-mapping shifts such as rotation and gain, linear calibration already restores success and ReGround matches it while lowering the unsafe rate; in several settings one-step inverse-dynamics transfer has a lower unsafe rate than ReGround but usually lower success. The monitor false-alarms on 2.1% of unshifted episodes.
