Dynamic Jumps of an Agile Bicycle Through Reinforcement Learning

Controlling bicycles during dynamic maneuvers, such as jumps, poses significant challenges due to the intricate demands of balance and stability. This study utilizes the Proximal Policy Optimization (PPO) algorithm to control both ground navigation and dynamic jumps. Stability on flat terrain is maintained through steering adjustments, while jumps are facilitated by vertical thrust. The proposed approach is validated through simulation, showing high success rates in jump execution, along with improved balance and adaptability. These results underscore the potential of reinforcement learning-based control systems in managing complex dynamic movements and contribute to the development of more robust motion control strategies for bicycles on both structured and unstructured terrains. The attached video is given in https:/lingi-ye.github.io/video/flybike.mp4.