Design and reinforcement learning-based locomotion gait planning for an 8-bar single-loop transformable robot

Abstract

Transformable robots with closed-chain mechanisms that exhibit underactuated and nonlinear dynamics pose a formidable challenge in gait planning. This paper introduces a two-stage reinforcement learning (RL) framework that bypasses the need for accurate system modeling to autonomously discover and refine locomotion gaits. We apply this method to an 8-bar single-loop transformable robot, whose kinematic analysis confirms its reconfigurability and variable degrees of freedom (DoF) nature. Our approach successfully generating two distinct and functional gaits: a stable peristaltic gait and a dynamic rolling gait that strategically leverages singular configurations. Extensive experiments on a physical prototype validate the gaits’ effectiveness across various terrains, with the robot achieving displacements of up to 870 mm at speeds of 8.7 mm/s. This work demonstrates a learning-based paradigm for planning complex locomotion in reconfigurable robotic systems.