Publikation
Federated Residual Reinforcement Learning for Collaborative Robot Skill Learning in Industry
Khalil Abuibaid; Vinit Vikas Hegiste; Tatjana Legler; Achim Wagner; Martin Ruskowski
In: 2025 3rd International Conference on Federated Learning Technologies and Applications (FLTA). International Conference on Federated Learning Technologies and Applications (FLTA-2025), Pages 530-536, IEEE, 2025.
Zusammenfassung
Collaborative robot learning has achieved significant progress in robotics for achieving sustainable adaptive controllers. However, to implement it in industry remains limited due to data-sharing restrictions and the need for robust, stable controller performance. This paper presents a Federated Residual Reinforcement Learning framework for collaborative robot skill learning without centralized data collection. The approach combines Residual Reinforcement Learning, based on Soft Actor-Critic, with a hybrid force–motion controller for adaptive manipulation, and Federated Learning for private and scalable model aggregation between clients via Federated Averaging. Two simulated UR5e robots are trained to learn optimal policies for peg-in-hole insertion with square and pentagon pegs, each featuring different friction coefficients. Experiments show that the global model outperforms individual client models in success rate, faster time-steps, reward maximization, and policy generalization. These results demonstrate Federated Residual Reinforcement Learning as a scalable, privacy-preserving, and sample-efficient solution for multi-robot skill learning.