FedDRMan distills vision-language-action policies between heterogeneous robots through federated learning, and gains 11.6 points over the best competing federated method.
Bringing together robots that do not manipulate the same objects or the same gestures usually means centralizing their demonstrations. Federated learning promises the opposite: improving a common policy without ever gathering data. However, the very dissimilar contributions of the different robots must not neutralize each other.
This is the problem addressed by FedDRMan, submitted to arXiv on September 26, 2026 by a team including Biprodip Pal, Kaushik Roy and Peyman Moghadam. The method distills vision-language-action policies into a low-ranking multimodal subspace and works on the distribution of actions, so as to preserve the specific representation of each robot. Compatible updates are grouped into clusters, each with a persistent server model, and then spectral rebalancing prevents the deletion of useful but minority directions.
On the LIBERO suite, with varying degrees of heterogeneity and participation rates, the method achieved an average success rate of 80.7%, 11.6 points higher than the best federated benchmark.
Read the full article on the original source
Go to the source →