Sim-to-Real: A Performance Comparison of PPO, TD3, and SAC Reinforcement Learning Algorithms for Quadruped Walking Gait Generation

The performance of the state-of-the-art Deep Reinforcement algorithms such as Proximal Policy Optimization, Twin Delayed Deep Deterministic Policy Gradient, and Soft Actor-Critic for generating a quadruped walking gait in a virtual environment was presented in previous research work titled “A Comparison of PPO, TD3, and SAC Reinforcement Algorithms for Quadruped Walking Gait Generation”. We demonstrated that the Soft Actor-Critic Reinforcement algorithm had the best performance generating the walking gait for a quadruped in certain instances of sensor configurations in the virtual environment. In this work, we present the performance analysis of the state-of-the-art Deep Reinforcement algorithms above for quadruped walking gait generation in a physical environment. The performance is determined in the physical environment by transfer learning augmented by real-time reinforcement learning for gait generation on a physical quadruped. The performance is analyzed on a quadruped equipped with a range of sensors such as position tracking using a stereo camera, contact sensing of each of the robot legs through force resistive sensors, and proprioceptive information of the robot body and legs using nine inertial measurement units. The performance comparison is presented using the metrics associated with the walking gait: average forward velocity (m/s), average forward velocity variance, average lateral velocity (m/s), average lateral velocity variance, and quaternion root mean square deviation. The strengths and weaknesses of each algorithm for the given task on the physical quadruped are discussed.