Reinforcement Learning-Based 2D Planar Locomotion Generation for Tendon-Driven Soft Robot

Abstract

Soft robots, with their high environmental adaptability, face challenges in control complexity. Simulation-based reinforcement learning often suffers from performance degradation when transferred to real systems. To address this, we applied reinforcement learning directly to an electrically tendon-driven soft robot to achieve leveled two-dimensional plane crawling. We developed a soft caterpillar robot with four motors enabling independent tendon control and twisting motions. Through 2.5 hours of reinforcement learning on the physical robot, it successfully reached a target 500 mm away in about 50 seconds while demonstrating various gaits. This approach enhances gait diversity and expands the robot’s potential for real-world applications, particularly in navigating uneven terrain.