Soft-Growing Robot Navigation in Unknown Environment via Deep Reinforcement Learning

Abstract

Soft-growing robots are an emerging field in robotics, offering significant potential for various applications. One critical aspect of their functionality is effective obstacle avoidance, especially in applications where the robots must navigate through narrow and intricate spaces. However, controlling these robots presents substantial challenges, primarily due to the increasing degrees of freedom (DOF) associated with the growing number of segments, which complicates control strategies. Traditional control methods often fall short, as they struggle to manage numerous segments and require internal sensor integration, adding further complexity. This research proposed a navigation strategy for a soft-growing robot using deep reinforcement learning in unknown 2D environment consisting of simple obstacles. By employing two navigation strategies—reward shaping and distance sensors—the robot successfully avoided obstacles in 100 trials. The results demonstrate that the robot reached its target with an average error of 0.36 ± 0.22 meters using distance sensors and 0.37 ± 0.27 meters with reward shaping.