Which simulation platform is used by DeepMind and others to handle inverse kinematics musculoskeletal simulation, etc., for reinforcement learning simulations and agents?
I thought they use Unity or Unreal but I assume that would be resource-heavy.
$\begingroup$
$\endgroup$
-
1$\begingroup$ I don't think there's one answer. For example, OpenAI developed their Gym for this purpose. $\endgroup$ – Oso Apr 14 '20 at 15:43
-
$\begingroup$ @Andy makes sense. So are these all custom developed? I wished someone could have had done some study of pro's and con's for these such as Gazebo, Cinder etc. I am quite eager to adapt the platform that is lightweight (not too many libraries) and highly controlable instead of having to deal with a framework. $\endgroup$ – gfdsal Apr 14 '20 at 16:15
-
$\begingroup$ I'm not sure, sorry, and a quick search didn't turn up much. What you're looking for is a survey of reinforcement learning toolkits, I think. $\endgroup$ – Oso Apr 14 '20 at 16:52
-
1$\begingroup$ Welcome to SE:AI! $\endgroup$ – DukeZhou♦ Apr 14 '20 at 22:17
-
$\begingroup$ @gfdsal You could try to ask this as a new question on our site. For example, you could ask "what are the pros and cons of simulators, such as Gazebo, Cinder, etc., for reinforcement learning tasks?" or something like that. Try to be specific and clear! $\endgroup$ – nbro♦ Apr 15 '20 at 21:21
$\begingroup$
$\endgroup$
DeepMind used MuJoCo (see also the related paper MuJoCo: A physics engine for model-based control) for the simulations, as they stated in section 3.1 of their paper Emergence of Locomotion Behaviours in Rich Environments (2017), which is the paper you should read to know more about their results related to those animations of skeletons that try to walk or jump (but do it weirdly).
-
$\begingroup$ This makes sense. I was told that Gazebo is one of the most used simulators in robotics as it has numerous builtin plugins for handing sensor data, noise. I am just confused why DeepMind and OpenAI and others develop their own version when there is something that is lot more versatile already exists. Gazebo even have solver for wind dynamics. I shall ready the paper you added to see key advantage of MuJoCo $\endgroup$ – gfdsal Apr 16 '20 at 12:32
-
1$\begingroup$ @gfdsal Honestly, I don't know right now. I could investigate that later, though. $\endgroup$ – nbro♦ Apr 16 '20 at 12:46
