We propose FedVLA, a federated learning framework for Vision-Language-Action models that enables cross-embodiment robot learning without centralizing sensitive data. Our method introduces an embodiment-aware aggregation strategy that accounts for heterogeneity across different robot morphologies. We evaluate on the Open X-Embodiment dataset across 6 robot platforms with ablation studies analyzing federation rounds, client heterogeneity, and communication efficiency. The framework addresses privacy concerns and institutional collaboration barriers while providing a complete implementation roadmap requiring approximately 500 GPU-hours for validation.
Key findings
Embodiment-aware aggregation weights client updates by morphology similarity, improving convergence across heterogeneous robots.
Federated learning enables training competitive VLA policies without sharing proprietary robot data across institutional boundaries.
Comprehensive evaluation protocol spans 6 distinct robot platforms from the Open X-Embodiment dataset with standardized metrics.
Systematic ablation studies quantify the impact of federation hyperparameters including local epochs and communication rounds.
Limitations & open questions
High computational requirements (500 GPU-hours) may limit accessibility for smaller research labs.
Effectiveness depends on sufficient embodiment diversity across federated clients to enable meaningful cross-embodiment generalization.
Regulatory compliance (GDPR/CCPA) adds complexity to federated deployment despite privacy-preserving architecture.