Locomotion Optimization of Photoresponsive Small-scale Robot: A Deep Reinforcement Learning Approach

Soft robots comprise of elastic and flexible structures, and actuatable soft materials are often used to provide stimuli-responses, remotely controlled with different kinds of external stimuli, which is beneficial for designing small-scale devices. Among different stimuli-responsive materials, liquid crystal networks (LCNs) have gained a significant amount of attention for soft small-scale robots in the past decade being stimulated and actuated by light, which is clean energy, able to transduce energy remotely, easily available and accessible to sophisticated control. One of the persistent challenges in photoresponsive robotics is to produce controllable autonomous locomotion behavior. In this Thesis, different types of photoresponsive soft robots were used to realize light-powered locomotion, and an artificial intelligence-based approach was developed for controlling the movement. A robot tracking system, including an automatic laser steering function, was built for efficient robotic feature detection and steering the laser beam automatically to desired locations. Another robot prototype, a swimmer robot, driven by the automatically steered laser beam, showed directional movements including some degree of uncertainty and randomness in their locomotion behavior. A novel approach is developed to deal with the challenges related to the locomotion of photoresponsive swimmer robots. Machine learning, particularly deep reinforcement learning method, was applied to develop a control policy for autonomous locomotion behavior. This method can learn from its experiences by interacting with the robot and its environment without explicit knowledge of the robot structure, constituent material, and robotic mechanics. Due to the requirement of a large number of experiences to correlate the goodness of behavior control, a simulator was developed, which mimicked the uncertain and random movement behavior of the swimmer robots. This approach effectively adapted the random movement behaviors and developed an optimal control policy to reach different destination points autonomously within a simulated environment. This work has successfully taken a step towards the autonomous locomotion control of soft photoresponsive robots