Learning Robust and Diverse Humanoid Jumping via Curriculum-Guided Whole-Body Coordination

Le résumé fourni par la source

Versatile jumping remains a significant challenge for humanoid robots, as it involves highly dynamic motions such as take-off, aerial flight, and landing impact absorption. These phases demand rapid force generation, precise posture regulation, and dynamic balance under underactuated conditions, placing stringent requirements on control strategies. This paper presents a multi-stage curriculum-assisted reinforcement learning frame-work, where the robot first acquires fundamental jumping skills through reference motion imitation, and subsequently generalizes to multi-target jumping tasks. A tailored reward mechanism promotes whole-body coordination, particularly focusing on torso orientation and arm movements, thereby improving both stability and versatility. Simulation results on the Unitree G1 humanoid robot show that the learned policy enables various jumps, including a forward jump of 1.8 m, lateral jump of 0.8 m, vertical jump of 0.4 m, and aerial yaw rotation of up to 115°. Under 0.1 s impulse perturbations up to 250 N, the robot maintains over 80% success rate in stable landings, indicating strong robustness and task generalization capabilities.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.