UC San Diego

This thesis develops a hierarchical control approach for hexapod robot navigation, consisting for a high-level vision-guided motion planning and a low-level kinematic gait generation. The low-level controller converts body-velocity commands into full eighteen-degree-of- freedom joint trajectories using inverse kinematics and a kinematics-based tripod gait generator. The high-level motion planner selects body-velocity commands from visual, depth, proprioceptive, and target-state inputs using reinforcement learning. A proximal policy optimization (PPO) baseline for body-velocity planning is first established. Two extensions are then proposed to improve the learning performance: a mixture-of-experts (MoE) architecture for multi-modal body-velocity command selection, and Replay-Enhanced Clipped Actor-Critic Policy Learning (RECAP) for improved sample efficiency. Route-level planning via a rapidly-exploring random tree with rewiring (RRT*) planner is also integrated to improve the safety and quality of the generated motion plans. The integrated approach, including RRT* planning, body-velocity generation via MoE and RECAP, and low-level gait control is demonstrated in simulation and on a physical hexapod robot across cluttered and moving-obstacle environments.