I'm looking for a reinforcement learning library that can be used for real-time robot control. What I first had in mind was ROS to describe the robot, Mujoco to simulate physics, and OpenAI gym to encapsulate RL. However, to my knowledge, this stack is only suitable when the robot is in simulation. What is the best way to implement and benchmark RL algorithms independent of the underlying control mechanism? Meaning, I would like to write the same algorithms, and just swap out simulated and real robot as needed, but the algos shouldn't know the difference. Thanks.