Data and benchmarks
Learning a Common Robot Policy Across Separate Laboratories
Federated VLA training can use demonstrations where they were collected, but its outcome depends on which parameters circulate, how sites differ, and whether personalization replaces the global model.
A group of robot laboratories can jointly adapt a shared vision-language-action policy by retaining demonstrations at their original sites while federating selected model parameters. The raw records remain local, while the collaboration operates on parts of the policy rather than pooling the demonstrations themselves.
The ForgeVLA proposal illustrates how this can work when sites hold vision-action pairs without language labels. Each participating client uses an embodied instruction classifier to map those pairs to a predefined instruction set, forming vision-language-action triplets for federated training without centralizing the raw data or requiring manual annotations. The available excerpt describes the framework but provides no comparative measurements.
The empirical study reports experiments with three modern pretrained vision-language-action policies. The study covered 40 simulated LIBERO manipulation tasks and six physical-robot tasks in two real-robot experiments, with demonstrations distributed across two and three sites, respectively. These findings are reported by the authors and are not independent confirmation.
Their experiments varied which policy parameters were federated, compared three aggregation algorithms, and tested distribution shift. The authors report that parameter scope had a greater effect than the choice among the aggregation algorithms. They also say federated fine-tuning could match centralized fine-tuning on heterogeneous data and remain at least as robust under distribution shift, although the abstract supplies no task-level measurements for those comparisons.
The physical-robot experiments reveal a limit: federated fine-tuning had difficulty matching centralized training when cross-site heterogeneity was stronger than simulation represented. Results from a simulated federation therefore do not establish that the same configuration will work across physical robots, environments, and locally collected demonstrations.
Personalization changes the objective. The study reports that allowing each site to federate one part of the policy while keeping another part local helped where pretraining was weak, but produced no usable global model. Laboratories seeking one deployable shared policy must therefore choose the federated parameter scope for that outcome and evaluate it under the differences between their sites; personalized local policies are not equivalent to a reusable global policy.