Teleoperation guided deep reinforcement learning robotics in a manufacturing environment

Project aim

To explore deep reinforcement learning for robotics in a manufacturing environment. The goal is to demonstrate that a robotic system is capable of solving manufacturing problems with the help of virtual reality (VR) hardware, imitation learning, and neural networks.

Project background

Robotics in a manufacturing context relies on hard-coded solutions using a teaching pendant or offline programming. These methods offer solutions to very specific tasks and fail when encountering slight changes in the environment. Advances in machine learning and specifically reinforcement learning solve this problem through trial and error. However, careful reward function engineering is needed, which is infeasible when applied to large-scale problems. This can be alleviated with virtual reality (VR) hardware. A demonstrator can be used to teach the robot a primitive action.

By teaching primitive skills through the use of imitation learning, a master controller can be trained to schedule primitive actions into a complex sequence that can be used to solve hard problems in a manufacturing context such as the automation of the hot-mounting press machine and the grinding and polishing machine. Additionally, it can be further taught to handle more problems because of its dynamicity.