Reinforcement and Imitation Learning for Diverse Visuomotor Skills

摘要: We propose a model-free deep reinforcement learning method that leverages small amount of demonstration data to assist agent. apply this approach robotic manipulation tasks and train end-to-end visuomotor policies map directly from RGB camera inputs joint velocities. demonstrate our can solve wide variety tasks, for which engineering scripted controller would be laborious. In experiments, imitation agent achieves significantly better performances than agents trained with or alone. also illustrate these policies, large visual dynamics variations, achieve preliminary successes in zero-shot sim2real transfer. A brief description work viewed https URL