Learning of Obstacle Avoidance with Redundant Manipulator by Hierarchical SOM

This paper proposes a method of obstacle avoidance motion generation for a redundant manipulator with a Self-Organizing Map (SOM) and reinforcement learning. To consider redundancy, two types of SOMs - a hand position map and a joint angle map - are combined. Multiple joint angles corresponding to the same hand position are memorized in the proposed map. Preserved redundant configuration information is used to generate motions based on tasks and situations, while resolving inverse kinematics problems with a redundant manipulator. The proposed map is applied to planning motion control using reinforcement learning in an unknown environment, where collision with obstacles is detected only directly by tactile sensing. The feasibility of the proposed framework was verified by simulation and experiments with an arm robot with force and a vision sensors.