Flow-based policies such as π0 are trained on a fixed set of robot embodiments, and transferring them to new ones remains an open problem. A new arXiv preprint proposes 'Embodiment-Classifier Guidance' to address this.
The paper's abstract says the goal is to adapt a frozen π0 to a new embodiment 'without additional demonstrations or parameter…' — the source text is truncated at that point, so the full mechanism is not described. The title indicates the method uses an embodiment classifier to guide the frozen policy.
If the approach works as described, it could let existing vision-language-action policies be reused across different robot hardware without collecting new demonstration data or retraining. Because this is a single preprint, the source offers no independent verification or comparison with other transfer methods.