RoboFL introduces a federated learning framework for vision‑language‑action and world‑action models. It addresses the problem of scarce, siloed, and heterogeneous robot interaction data by letting each client train a LoRA adapter on its own task family. Instead of averaging these adapters naively, RoboFL installs them as expert branches in a server‑side mixture‑of‑experts (MoE) model. Server routers learn token assignments over these prior‑informed branches, while two innovations—FARD and PCEA—align routing across model paths and convert expert updates into a compact global adapter for redistribution.
Experiments on the RoboTwin 2.0 and RLBench simulators and a real‑world Franka robot arm show that RoboFL outperforms a centralized parameter‑efficient fine‑tuning baseline (InternVLA‑A1) by 12.23% on the Franka arm. It also cuts per‑round client communication by up to 86.81% compared to MoE‑based federated VLA baselines. The method is interesting because it preserves local specialization while building a shared policy, overcoming a key challenge in federated robot learning where clients have incompatible task distributions.