Recurrent Neural Network based Partially Observed Feedback Control of Musculoskeletal Robots

The musculoskeletal robot has become a promising research direction. However, the control problem of the musculoskeletal robot still limits its application. Especially in the real-world application, only some incomplete and imprecise feedback states could be obtained due to the limitation of sensors. Therefore, this paper proposes a recurrent neural network (RNN) based partially observed feedback control method of musculoskeletal robots. The RNN has the ability of working memory and can realize muscle control through implicitly inferring sufficient states from partially observed states. It can also generate muscle excitations with synergies to guarantee great generalization. The effectiveness of the proposed method is verified on a simulated musculoskeletal system. Compared with other deep reinforcement learning method, the proposed method achieves similar performance under sufficient feedback states and better performance under partially observed feedback states.