Recovery from Injury: Learning Bipedal Jumping Skills with a Motor Output Torque Limit Curriculum.

Bipedal robots can walk and run on different terrains and show great capacity in fast-moving. However, it’s still a daunting challenge for them to achieve highly dynamic whole-body motions such as jumping. In this paper, we propose a method to learn high jump skills for humanoid robots, and the effectiveness of the method is verified through simulations on the Ranger Max humanoid robot model. Both 2D and 3D jumping locomotion for the one-legged and two-legged Ranger Max robots are generated naturally and stably with different scales of maximum motor output torque limit. A curriculum learning strategy inspired by the idea of “recovery from injury” is proposed to make the learning of the high jump more efficient for “weaker” robots, which is confirmed by the simulation results.