Tangle- and contact-free path planning for a tethered mobile robot using deep reinforcement learning

Abstract

Read online

This paper presents a tangle- and contact-free path planning (TCFPP) for a mobile robot attached to a base station with a finite-length cable. This type of robot, called a tethered mobile robot, can endure long-time exploration with a continuous power supply and stable communication via its cable. However, the robot faces potential hazards that endanger its operation such as cable snagging on and cable entanglement with obstacles and the robot. To address these challenges, our approach incorporates homotopy-aware path planning into deep reinforcement learning. The proposed reward design in the learning problem penalizes the cable-obstacle and cable-robot contacts and encourages the robot to follow the homotopy-aware path toward a goal. We consider two distinct scenarios for the initial cable configuration: 1) the robot pulls the cable sequentially from the base while heading for the goal, and 2) the robot moves to the goal starting from a state where the cable has already been partially deployed. The proposed method is compared with naive approaches in terms of contact avoidance and path similarity. Simulation results revealed that the robot can successfully find a contact-minimized path under the guidance of the reference path in both scenarios.

Keywords

• tethered mobile robot
• path planning
• homotopy class
• reinforcement learning
• deep Q-network