Design and Control of Continuous Jumping Gaits for Humanoid Robots Based on Motion Function and Reinforcement Learning

Continuous jumping of humanoid robots is a challenging problem that requires appropriate gait and force control for the robot to achieve. Reinforcement learning (RL) has been applied to the motion control of humanoid robots and has achieved significant progress. However, traditional RL control methods primarily rely on designing reward signals to generate human-like movements, and it requires developing complex reward signals for continuous jumping, resulting in limited actions and suboptimal jumping performance. To address this issue, we have designed motion functions for humanoid continuous jumping and incorporated them into the RL framework to obtain desired human-like continuous jumping actions. Furthermore, two different jumping gaits are designed to investigate the jumping effects produced by different jumping functions: bent-knee jumping (BJ) and straight-knee jumping (SJ). BJ is a typical jumping style where the legs flex during the jumping process to accumulate energy, enabling higher explosive power. SJ is more suitable for tasks requiring quick responses, as the legs remain upright during the jump, providing faster reaction speed. The performance in continuous jumping tasks has been evaluated by comparing the effects of training with different jumping functions. This research has particular guidance for improving the jumping ability of robots and enhancing the stability and efficiency of humanoid continuous jumping.