Dynamic terrain gait control of biped robots based on environmental perception and reinforcement learning

For bipedal robots performing tasks such as rescue and resource exploration in the wild, traditional pre-programmed gaits have poor adaptability and are prone to instability on dynamic and unstructured terrains like grasslands, slopes, and gravel roads. This paper proposes a dynamic terrain gait adaptive control method based on environmental perception and reinforcement learning. By integrating visual terrain preview from RGB-D cameras, contact feedback from six-axis force sensors on the soles, and the robot's body posture estimation from IMUs, a real-time terrain feature and robot state perception model is constructed. Based on this, a hierarchical control framework is designed: the upper layer uses a gait generator based on deep reinforcement learning to generate adaptive sole trajectories and trunk postures online according to the perceived terrain features. the lower layer employs an impedance controller based on the full-body dynamics model to achieve compliant contact between the feet and the ground and resist disturbances. Experimental results show that this approach enables the robot to maintain stable walking on various dynamic terrains, with the terrain passage success rate increasing by over 30% and posture fluctuations reducing by approximately 40% compared to the traditional ZMP gait, verifying its strong robustness and adaptability in unknown complex environments.