Robot arm motion planning for safe execution of user commands against dynamic obstacles

Abstract

Remote manipulation is to control a robot from a user command and has been primarily used to perform dangerous or difficult tasks for humans in special places such as nuclear power plants or medical facilities. Recently, the scope of remote manipulation has expanded to environments around us (e.g., homes and convenience stores), and studies have been conducted to perform tasks through robots without directly going to those places. Since environments around us have many different and dynamic obstacles, it is necessary to perform user commands safely from these obstacles. Obstacle avoidance requires high computational costs, whereas users expect immediate execution of commands. Therefore, it is an important issue to find collision-free joint configurations that follow user commands while reducing the amount of computation for obstacle avoidance. In this dissertation, we present robot arm motion planning approaches for safe execution of user commands against various and dynamic obstacles. This dissertation also targets a redundant manipulator, which has various joint configurations for a user command given to the manipulator's end-effector. Accordingly, this dissertation deals with finding a collision-free joint configuration among various solutions for a user command and adjusting a risky command due to the user's unpredictability or carelessness to be safe. Specifically, we propose 1) a trajectory optimization method for generating joint configurations that follow a given end-effector path avoiding obstacles. We also propose 2) real-time collision-free inverse kinematics based on deep learning to perform consecutive user's real-time commands in dynamic environments. Lastly, we propose 3) a reinforcement learning-based user command adjustment method to adjust risky commands caused by the user's carelessness to be safe. We showed that our proposed methods increase safety from various and dynamic obstacles through experiments using the real robot and sensor data.