Neural networks for selecting actions to be performed by robot agents

摘要

The system includes a neural network system implemented by one or more computers. The neural network system receives and responds to observations that characterize the current state of the real-world environment that is interacted with and interacts with the robot agent to perform robot tasks. And processing the observations to generate a policy output that defines the actions to be performed by the robot agent. A neural network system includes a deep neural network (DNN) sequence, where the DNN sequence (i) performs a simulated version of a robot task, a simulated version of a robot agent, and a real world Simulation trained DNN trained on interaction with a simulated version of the environment, and (ii) receive observations and process observations to generate policy output And a first robot-trained DNN configured.