Navigating a robotic arm to reach a target position with its end-effector under constrained environment is always a challenging task especially when the environment is only partially known. A similar but practical example for this situation is to let a multi-links robot, PIRATE (a pipe-inspection robot developed in the university of Twente), move autonomously in industrial pipes which have turns, splits or sections with varying diameters, while the robot is only able to detect part of the pipes by using its laser sensors or camera and has to make approximations on the knowledge of the environment.
One possible solution for the situation above is to use a learning-based motion planner for the robot by applying Reinforcement Learning (RL). RL is an machine learning technique which tries to find an optimal policy for an agent to achieve certain tasks by continuously interacting with the environment and improving the policy with gained experience.
Therefore, this project will focus on designing a robust, high-level controller for a multi-links robotic arm by using Reinforcement Learning in order to make the arm navigate itself under different constrained environments.