Gait transition strategy for hexapod robots based on reinforcement learning

Purpose – Hexapod robots possess a diverse range of gaits, and dynamic transitions between different gaits are of great significance for enhancing the overall motion performance of the robot. Therefore, to address the problem of dynamic optimal gait transition, a gait transition strategy for hexapod robots based on reinforcement learning is proposed. The purpose of this study is to achieve a dynamic, stable, and fast gait transition for hexapod robots from an initial state to a target gait. Design/methodology/approach – First, an extended discretization method is introduced based on the robot’s foot motion states to define a stable and reachable motion state space. Second, the kinematic stability margin of each robot state is computed, and a gait transition strategy that satisfies six motion constraints is designed. On this basis, a Markov decision process is incorporated, and obtaining a near-optimal gait switching list within the discrete state-action space of the model and the specified reward design scope so as to achieve gait switching with an optimal balance between switching speed and motion stability. Findings – Experimental validation is conducted using the BIT Nahex wheel-legged robot, the expanded discrete-point processing method for the selectable motion space can effectively mitigate the impact of motion constraints in gait transition strategies. This approach enables the robot’s feet to achieve both rapid gait switching and optimal stability margins, realizing dynamically free gait transitions with comprehensively optimized motion performance under specific model conditions. Originality/value – A gait transition strategy for hexapod robots based on reinforcement learning is proposed and verified. Experiments verify that this switching strategy exhibits excellent adaptability to the dynamic switching from different initial states to the target gait within the stable and reachable state space, and the effectiveness of each component of the method is verified through comparative experiments.