Towards Dynamic Obstacle Avoidance for Robot Manipulators with Deep Reinforcement Learning
Publikation aus Robotics
Friedemann Zindler, Matteo Lucchi, Lucas Wohlhart, Horst Pichler, Michael Hofbaur
Advances in Service and Industrial Robotics. RAAD 2022. Mechanisms and Machine Science, vol 120. Springer, Cham. , 4/2022
We use reinforcement learning (RL) to demonstrate an easily reproducible setup to learn dynamic obstacle avoidance for a robotic arm based on sensory input as it follows a pre-planned trajectory. Training takes place exclusively in a simulation environment with random obstacle movements around the robot. We show that training dynamic obstacle avoidance in simulation translates well to the real environment with a UR5 manipulator and yields similar performance and success without further tuning of the learned policy. This is a step towards learning general skills needed to enable robots to operate in dynamic environments shared with humans. Source code, data and application videos are available at: https://www.robogym.net.
Keywords: Machine learning for robot control, Collision avoidance, Transfer learning