Research on Jumping Gait Motion Control of Quadrupedal Robots Based on Reinforcement Learning

Abstract

High-dynamic quadruped jumping over steps and gap-like support interruptions remains difficult when external terrain perception is unavailable or unreliable. This paper studies a proprioception-only reinforcement-learning policy for bound jumping. The actor is initialized by a flat-ground bound strategy to obtain a stable fore-hind jumping rhythm and is then continuously optimized through curriculum learning in step and predefined gap-like support-interruption environments. The deployed actor uses velocity commands, joint states, inertial information, gait phase, foot-contact history and short-term action history, whereas terrain-specific privileged information and support-state annotations are used only for reward computation and critic learning during training. In simulation, the policy is evaluated at a 1.2 m/s forward command. It achieves success rates of 85.0%, 78.0% and 72.0% on 0.08 m, 0.12 m and 0.16 m steps, respectively, and 80.0%, 69.0% and 41.0% on 0.15 m, 0.25 m and 0.35 m gap-like support interruptions, respectively, outperforming the flat-ground-only and no-phase-contact variants in traversal success and forward-speed recovery. These results show that gait phase observation and multiple reward shaping can improve high dynamic jumping under limited perception. More importantly, this study demonstrates the potential of proprioceptive reinforcement learning only in perceptually constrained agile motion, especially when external terrain perception is unavailable, degraded or unreliable. The framework provides a lightweight and deployable control direction for quadruped robot operation on discontinuous terrain without relying on online topographic map.