Reinforcement learning-based obstacle avoidance for continuum robots

Abstract

This work presents a reinforcement learning framework for controlling a planar threesection continuum robot in environments with static obstacles. Assuming constant curvature for each section, the robot is trained to navigate toward a fixed goal while avoiding collisions with multiple static objects. A custom simulation environment was developed to support three levels of scenario difficulty, easy, medium and hard, each with varying obstacle density and placement. The learning process is driven by the Deep Deterministic Policy Gradient (DDPG) algorithm, which enables smooth and continuous curvature control. Careful attention was paid to the design of the reward function and the network architecture, both of which were critical to achieving stable and reliable policy learning. Performance was evaluated across multiple runs, revealing that the agent successfully generalized its behavior across scenarios of increasing complexity. The proposed framework demonstrates the potential of reinforcement learning as a viable approach to safe and adaptive control in continuum robotic systems, with promising implications for applications such as medical navigation, search and rescue, and inspection in confined environments.