A system for dexterous robot manipulation includes a first robot configured to perform real simulations. The system also includes at least one computer configured to perform a set of virtual simulations including a robot model, develop a first policy for robot maneuvering based on the set of virtual simulations, and record a trajectory of the robot model during the set of virtual simulations. The at least one computer is also configured to perform a set of real simulations including the first robot, where the first robot mimics the recorded trajectory of the robot model, and develop a second policy for robot maneuvering based on the set of real simulations. The at least once computer is also configured to deploy the second policy to at least one of the first robot and a second robot to perform dexterous manipulation.