- Abstract: We propose a general deep reinforcement learning method and apply it to robot manipulation tasks. Our approach leverages demonstration data to assist a reinforcement learning agent in learning to solve a wide range of tasks, mainly previously unsolved. We train visuomotor policies end-to-end to learn a direct mapping from RGB camera inputs to joint velocities. Our experiments indicate that our reinforcement and imitation approach can solve contact-rich robot manipulation tasks that neither the state-of-the-art reinforcement nor imitation learning method can solve alone. We also illustrate that these policies achieved zero-shot sim2real transfer by training with large visual and dynamics variations.
- TL;DR: combine reinforcement learning and imitation learning to solve complex robot manipulation tasks from pixels
- Keywords: reinforcement learning, imitation learning, robotics, visuomotor skills