Data collection where the robot executes actions from its current policy to generate training trajectories.