[2G6-OS-47c] OS-47

5:45 PM - 6:00 PM JST(8:45 AM - 9:00 AM UTC)

[2G6-OS-47c-02] Development of Physics Simulator for Reinforcement Learning of Autonomous Maintenance and Inspection Robots

〇Hayato Kojima¹, Yuki Tanaka¹, Seiichiro Katsura¹ (1. Keio University)

Keywords:

Reinforcement Learning,Robotics,Simulation

In recent years, the declining working population due to a low birthrate and aging society, coupled with the advancement of robots utilizing artificial intelligence, has heightened the importance of employing robots for the maintenance and inspection of infrastructure facilities. The tasks such as verifying the operation of breakers and switches on control panels, replacing deteriorated components, and inserting parts into each other necessitates the establishment of complex mechanical contact and grasping with the target components. This renders automation difficult using conventional methods. This research describes the development of a physics simulator specialized for inspection and a method for acquiring actions using reinforcement learning to automate such tasks. First, a simulation environment was constructed using a physics simulation engine to reproduce the components and physical characteristics inside the control panel. Subsequently, reinforcement learning was executed within this environment to acquire the task. The integration of pre-training in simulation with human correction operations is anticipated to facilitate autonomous execution of complex inspection and replacement tasks, thus reducing the burden on the equipment and the learning time required.