Implementation and Optimization of Robotic Jumping Behavior in a Virtual Simulation Environment Using Reinforcement Learning

This paper presents the implementation and optimization of bipedal jumping behavior for the Joyson Robot's Jarvis 1st generation robot in a virtual simulation environment using reinforcement learning. Leveraging the Isaac Gym platform, we employ the Proximal Policy Optimization (PPO) algorithm to train the robot to autonomously perform stable jumping in response to external forces. The developed jumping gait exhibits the following key characteristics: (1) Fast Response: The jumping gait can quickly generate significant force to rapidly counteract external disturbances. (2) Large Dynamic Range: The robot demonstrates the ability to handle larger external disturbances while maintaining stability in unstable environments. (3) High Energy Efficiency at High Speeds: The energy efficiency of the jumping gait improves during high-speed motion, making it more suitable for long-distance rapid movement. These results highlight the potential of reinforcement learning in enhancing the adaptability and performance of bipedal robots in dynamic and challenging environments.