← Paper Digest

Right Goal, Wrong Body: Re-grounding Frozen VLAs under Actuation Shift

Umer Siddique, Peilang Li, Yongcan Cao

November 2026

Simulated LIBERO rollouts under an actuation shift, side by side for the frozen VLA, linear calibration, intent map and ReGround.

TL;DR ReGround keeps a VLA frozen and corrects its actions at test time when the robot's actuation changes (latency, controller gain, action-frame rotation), using a dynamics model fitted on about a minute of motor babbling. On eight simulated LIBERO tasks it raises π0.5's success under 400 ms latency from 0.16 to 0.90.

What they do

A VLA is trained on one mapping from its action commands to robot motion. Remounting, wear, gain changes or added latency break that mapping (actuation shift) even though the task and camera images are unchanged. ReGround keeps the VLA frozen and, at inference time, rewrites its actions so the deployed robot follows the motion the VLA intended. It needs no rewards, demonstrations or retraining.

How it works

  • Intent read-out. A forward model of the training body, fitted once on nominal babbling, rolls out the VLA’s 50-step action chunk to get the intended end-effector path.
  • Deployed-body model. Ornstein-Uhlenbeck motor babbling on the deployed robot fits an ensemble forward model and estimates the actuation delay and a linear action map.
  • Trust-region tracking. A receding-horizon CEM solver tracks the intended path in the deployed-body model while a quadratic penalty keeps actions near the VLA’s re-timed, re-mapped chunk.
  • Shift monitor. A one-sided CUSUM test on the training-body model’s residual turns ReGround on only after a shift is detected.

Results

  • Shifts break the frozen π0.5: validation success falls from 0.99 to 0.19 at 400 ms latency and to 0.00 at 90° rotation, and unsafe episodes rise from 0.09 to as much as 0.82.
  • On π0.5 across eight LIBERO tasks, ReGround reaches 0.90-0.98 test success in all five shift settings, e.g. 0.16 to 0.90 at 400 ms latency (+0.74 over the best baseline) and 0.97 on the composite shift versus 0.92 for the analytic inverse.
  • Where linear calibration already restores success (rotation 90°, gain ×3), ReGround matches it while cutting unsafe episodes from 0.76 to 0.29 and from 1.00 to 0.77.
  • The monitor catches every shifted episode with 2.1% false alarms; 1k babbling transitions (50 s) give 0.98 success at 250 ms latency, where the intent map needs 10k.

Limitations

The paper has no limitations section. All experiments are in simulation (LIBERO in MuJoCo/robosuite) with synthetic shifts. The one-step intent-map baseline is safer in several settings, mostly at lower success, and the unsafe rate stays at 0.77 under gain ×3.

Videos

The same four-way comparison in a LIBERO kitchen scene.
Linear calibration completes the task but knocks an object over, while ReGround completes it cleanly.

Figures

From Figure 1: A paired episode at 400 ms latency, the frozen π0.5 (top) against ReGround (bottom) at 2.5, 4.5 and 6....
From Figure 1: A paired episode at 400 ms latency, the frozen π0.5 (top) against ReGround (bottom) at 2.5, 4.5 and 6.4 s.
From Figure 1: ReGround reads the intended path out of the frozen VLA's action chunk with a training-body model, re-g...
From Figure 1: ReGround reads the intended path out of the frozen VLA's action chunk with a training-body model, re-grounds it with MPC in a deployed-body model fitted on 50 s of motor babbling, and passes VLA actions through until a shift monitor raises an alarm.
From Figure 3: Test success under each shift for the frozen VLA (grey), linear calibration (red), intent map (pink), ...
From Figure 3: Test success under each shift for the frozen VLA (grey), linear calibration (red), intent map (pink), ReGround (blue) and the analytic inverse oracle (hatched).
From Figure 3: Success at 250 ms latency against the number of babbling transitions (1k is about 50 s of robot time),...
From Figure 3: Success at 250 ms latency against the number of babbling transitions (1k is about 50 s of robot time), for the intent map and ReGround.
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 $π_{0.5}$'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/.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.