Learning Multi-mode Musculoskeletal Motion for Natural Walking and Running Gaits

Traditional musculoskeletal control policies primarily focus on individual gait patterns, with limited success in developing a single model capable of handling multiple locomotion modes, such as walking and running. While prior work has explored learning different gaits within a shared framework, existing approaches typically train independent models rather than a unified, adaptable policy. In this paper, we propose a multi-mode controller learning framework that leverages Feature-wise Linear Modulation (FiLM) to integrate mode-specific dynamics into a unified controller. By incorporating FiLM layers into the control policy, we enable efficient adaptation between locomotion modes while maintaining network simplicity. To further enhance generalizability and robustness, we incorporate skeleton scaling and muscle condition randomization, which broaden the model’s state-space exploration and significantly improve transition stability. Additionally, we introduce a novel reward design that regulates stride characteristics, facilitating the learning of natural walking and running behaviors across varying speeds. Our results show that the proposed framework produces coordinated, natural-looking gaits and smooth transitions between modes, demonstrating the effectiveness of FiLM in unified musculoskeletal locomotion control.