Adaptive Leg Motion Planning Method for Spherical Multiretractable Legged Robots Using Deep Reinforcement Learning

Abstract

Spherical multi-retractable legged robot possess the capability to move through intricate terrains and execute omnidirectional locomotion. Nonetheless, the current state of motion planning for this class of robot lacks a comprehensive analysis of individual leg dynamics. To address this gap, this study introduces an innovative leg motion planning framework centered around two pivotal critical angles. These angles accurately delineate leg extension and retraction actions. To ascertain the optimal combination of these critical angles, deep reinforcement learning algorithms are employed. This facilitates the robot’s acquisition of adaptive decision-making skills, enabling it to select the most suitable critical angles based on its own state and predefined trajectories. Furthermore, to mitigate the difference between simulation and real-world implementation, sim-to-real techniques are introduced. The proposed methodology is validated through simulation experiments on the PyBullet platform, encompassing various predefined trajectory types such as linear, curved, and circular. Ultimately, the effectiveness of the proposed method is corroborated through empirical validation via prototype experiments.