Multi-Goal Reinforcement Learning: Challenging Robotics Environments and\n Request for Research

The purpose of this technical report is two-fold. First of all, it introduces\na suite of challenging continuous control tasks (integrated with OpenAI Gym)\nbased on currently existing robotics hardware. The tasks include pushing,\nsliding and pick & place with a Fetch robotic arm as well as in-hand object\nmanipulation with a Shadow Dexterous Hand. All tasks have sparse binary rewards\nand follow a Multi-Goal Reinforcement Learning (RL) framework in which an agent\nis told what to do using an additional input.\n The second part of the paper presents a set of concrete research ideas for\nimproving RL algorithms, most of which are related to Multi-Goal RL and\nHindsight Experience Replay.\n