Multi-goal RL benchmark paper includes in-hand manipulation with Shadow Dexterous Hand

Evidence notes

• The report introduces OpenAI Gym continuous-control environments based on existing robotics hardware, including Fetch arm tasks and in-hand manipulation with a Shadow Dexterous Hand.
• The Shadow hand environment uses sparse binary rewards and a multi-goal reinforcement-learning setup in which the desired manipulation goal is supplied as an additional input.
• The report also proposes research directions for multi-goal reinforcement learning and Hindsight Experience Replay.