Ingredients for Robotics Research

today we're releasing new gym environments and baselines for goal-based robotics tasks the tasks include a shadow hand robot that manipulates an object the hand is flexible enough to handle a number of different items we also added new environments for the fetch robot where it learns to nudge a block slide a puck as well as grasp a block and lift it up along with the environments were also updating baselines our library of reinforcement learning code one of the algorithms we use to train agents on the new tasks is hindsight experienced replay or her for short her allows an agent to learn from failures for example here fetch is trying to slide a puck towards a goal unfortunately in this first attempt we miss the goal usually we would now just try again and hope for a better result however while the puck did not end up at the goal it did land somewhere else with her instead of the original goal we pretend that we in fact wanted to slide the puck to where it landed if our goal had been here this attempt would have been successful over time this helps us learn how to achieve all possible goals the new environments are available in gym today and hindsight experience replay is available in baselines you