Embodiment-Classifier Guidance: Transferring Frozen VLAs Across Robot Dynamics without Demonstrations
Embodiment-Classifier Guidance: Transferring Frozen VLAs Across Robot Dynamics without Demonstrations
This research proposes a method to adapt frozen vision-language agents (VLAs) to new robot embodiments without additional demonstrations. By training a lightweight embodiment classifier to distinguish successful action chunks on source and target bodies, the method nudges the denoising steps of the original policy with the classifier's gradient, improving success rates on transfer tasks. The approach shows promise in simulation but has limitations when data is insufficient or motions are infeasible.
This paper introduces a method to transfer frozen vision-language agents (VLAs) to new robot embodiments without additional demonstrations, using a lightweight embodiment classifier to guide the transfer process. The approach shows promise in simulation, improving success rates on transfer tasks, though it has limitations when the classifier lacks sufficient data or favors infeasible motions.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- Environment
- SimulationOpen source S4
Reported numbers
number of tasks
12
View original evidence
aises $\pi_0$'s average success on the target bodies from 72% to 80% across 12 suite-controller pairs, with a 31% relative reduction in failures (11 of 12 gain; 5 significant after Holm correction). In simulation, guidance is data-efficient, needing few target-body rollouts per task, which makes it promising for sim-to-real transfer, where rollouts are costly. Guidance is mixed on weak wrists and harmful on a locked
Open source S4number of successes
11
View original evidence
aises $\pi_0$'s average success on the target bodies from 72% to 80% across 12 suite-controller pairs, with a 31% relative reduction in failures (11 of 12 gain; 5 significant after Holm correction). In simulation, guidance is data-efficient, needing few target-body rollouts per task, which makes it promising for sim-to-real transfer, where rollouts are costly. Guidance is mixed on weak wrists and harmful on a locked
Open source S4number of significant results
5
View original evidence
aises $\pi_0$'s average success on the target bodies from 72% to 80% across 12 suite-controller pairs, with a 31% relative reduction in failures (11 of 12 gain; 5 significant after Holm correction). In simulation, guidance is data-efficient, needing few target-body rollouts per task, which makes it promising for sim-to-real transfer, where rollouts are costly. Guidance is mixed on weak wrists and harmful on a locked
Open source S4
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
aises $\pi_0$'s average success on the target bodies from 72% to 80% across 12 suite-controller pairs, with a 31% relative reduction in failures (11 of 12 gain; 5 significant after Holm correction). In simulation, guidance is data-efficient, needing few target-body rollouts per task, which makes it promising for sim-to-real transfer, where rollouts are costly. Guidance is mixed on weak wrists and harmful on a locked
Open source S4
wrist, when the classifier lacks data or favors infeasible motions. Collecting more target-body data with guided $\pi_0$ can fill such gaps: on one weak-wrist arm, guided $\pi_0$ solved a task on which unguided $\pi_0$ had zero successes in the original data, and retraining the classifier on the guided successes raised success from 52.5% to 70.7%.
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org