Embodiment-Classifier Guidance: Transferring Frozen VLAs Across Robot Dynamics without Demonstrations

Summary

Flow-based policies such as $\pi_0$ are trained on a fixed set of embodiments, and transferring them to new ones is an open problem.

Abstract Summary

Flow-based policies such as $\pi_0$ are trained on a fixed set of embodiments, and transferring them to new ones is an open problem. To adapt a frozen $\pi_0$ to a new embodiment without additional demonstrations or parameter updates, we train a lightweight embodiment classifier to distinguish successful action chunks on the source body from those on the target body, using a few calibration tasks. On transfer tasks that $\pi_0$ saw on its source body but the classifier was never trained on, we nudge each denoising step of $\pi_0$ with the classifier's gradient, which estimates the direction from the distribution of successful actions on the source body toward that on the target body. On the LIBERO simulation benchmark with four modified controllers, embodiment-classifier guidance raises $\pi_0$'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 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%.

Key Points

Flow-based policies such as $\pi_0$ are trained on a fixed set of embodiments, and transferring them to new ones is an open problem.
To adapt a frozen $\pi_0$ to a new embodiment without additional demonstrations or parameter updates, we train a lightweight embodiment classifier to distinguish successful action chunks on the source body from those on the target body, using a few calibration tasks.
On transfer tasks that $\pi_0$ saw on its source body but the classifier was never trained on, we nudge each denoising step of $\pi_0$ with the classifier's gradient, which estimates the direction from the distribution of successful actions on the source body toward that on the target body.
On the LIBERO simulation benchmark with four modified controllers, embodiment-classifier guidance raises $\pi_0$'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.

Embodiment-Classifier Guidance

Setting. $\pi_{0}$ maps an observation $o$ (two camera images, robot state, instruction) to an action chunk $a\in\mathbb{R}^{50\times 7}$.

It is a flow-matching model [ 17 ] : with the noised chunk $x_{t}=t\,\varepsilon+(1-t)\,a$ and flow time $t\in[0,1]$, a network $v(x_{t},t,o)$ predicts the velocity $\varepsilon-a$, and sampling takes 10 Euler steps from $x_{1}=\varepsilon$, after which the robot executes 5 actions and $\pi_{0}$ plans again.

Embodiment-Classifier Guidance — Embodiment-Classifier Guidance: Transferring Frozen VLAs Across Robot Dynamics without Demonstrations — Figure 1: Embodiment-classifier guidance steers a frozen $\pi_{0}$ toward plans that succeed on a target body. $\pi_{0}$, trained on a Panda, runs on a KUKA arm with 25% wrist strength. Middle (schematic): the classifier nudges every denoising step. Right: a selected scene where guidance succeeds and unguided $\pi_{0}$ fails.

Experimental Setup

Policy and tasks. We use the frozen public $\pi_{0}$ LIBERO checkpoint [ 2 ] in LIBERO [ 18 ] (robosuite [ 28 ] ), a simulated benchmark of tabletop manipulation tasks in four suites that vary the objects (Object), spatial layouts (Spatial) and goals (Goal), plus long-horizon tasks (LIBERO-10). Each suite has four calibration and four transfer tasks (App. A ), all of which $\pi_{0}$ saw on its source body.

Results

Across the four suites, guidance improves 11 of the 12 suite–body pairs with a changed controller, 7 with a CI above zero (Fig. 2 A, B); 5 remain significant after a Holm correction over the 12 pairs. Pooled over the 12 pairs, success rises from 72% to 80% (1,718 to 1,931 of 2,400 episodes), a 31% relative reduction in failures.

On LIBERO-Object, Ctrl(0.5, 0.3 s) rises from 73% to 93% (+20 points [+13.5, +27]); two more classifier seeds give 93.5% and 91%. The exception is LIBERO-10, where Ctrl(0.5, 0.3 s) drops by 5.5 points [$-13$, +2].

Results — Embodiment-Classifier Guidance: Transferring Frozen VLAs Across Robot Dynamics without Demonstrations — Figure 3: Toy illustration with made-up 2-D data (not robot data). Each point is a 2-D “action”. The source body’s actions form two clusters (dashed circles); the target body succeeds only in the shaded region. Lines show 60 samples moving from noise ($t=1$) to actions ($t=0$) under the exact flow of the source-body data. Left: the frozen flow puts 48% of samples in the shaded region. Right: adding a capped classifier gradient (here a fixed linear logit that prefers larger $x_{1}$) puts 63% there. Guidance mostly moves samples toward the source-body cluster that also works on the target body.

Sources and Demonstrations

Figures are reproduced from Han et al. See the full paper for experimental details and the project page, when the paper links one, for demonstrations and videos.