Deep reinforcement learning system for robotic manipulation

摘要

A non-transitory computer-readable storage medium having instructions stored thereon, the instructions, when executed by one or more processors, causing the one or more processors to perform steps comprising: while performing multiple episodes by each of a plurality of robots Each of the episodes is an examination of the performance of a neural network policy task that represents a reinforcement learning policy for the task: storing instances of robot experience data in a buffer generated by the robots during the episodes, each of the instances the robot experience data is generated during a corresponding one of the episodes and is generated, at least in part, at a corresponding output corresponding to the neural network using the neural policy network with corresponding policy parameters for the neural network iteratively generating updated policy parameters of the neural policy network, each of the iterations of iteratively generating comprising generating the updated policy parameters using a group of one or more instances of the robot experience data in the buffer during the iteration; andby each of the robots in conjunction with a start of each of a plurality of episodes performed by the robot, updating the neural policy network to be used by the robot in the episode, wherein updating the neural policy network calls for the use of the updated policy parameters recent iteration of iteratively generating the updated policy parameters.