Hierarchical Control Architecture for Quadruped Robot Gait Based on Hopf-CPG Parameterized Action Space with Reinforcement Learning

Abstract

To address the problems of high-dimensional action space in end-to-end reinforcement learning and insufficient adaptive capability of traditional CPG parameters in quadruped robot locomotion control, this paper proposes a hierarchical control architecture combining a Hopf-CPG parameterized action space with reinforcement learning. The architecture consists of three tiers: a learned SAC policy (25 Hz) that outputs 4-dimensional CPG parameter adjustments (step frequency, amplitude, swing height, stride length), deterministic Hopf-CPG oscillators with foot trajectory mapping and inverse kinematics, and a PD servo controller (1000 Hz). A sensory feedback modulation mechanism utilizing foot contact and body posture feedback online corrects oscillator phase and amplitude at 100 Hz, independent of the learned policy. The cost of transport (CoT) is defined as the ratio of total mechanical power to body weight times velocity. Simulation experiments on flat ground, slopes, low-friction, and uneven terrains demonstrate excellent gait stability and energy efficiency, with sensory feedback reducing disturbance recovery time by 56%. All results are reported as mean ± standard deviation over 20 runs with 5 random seeds.