pybullet-gym
pybullet-gym copied to clipboard
OpenAI Gym Robotics environments implementation
Hey so I want to use some of the Fetch robot environments in OpenAI but, for some weird reason, they continue to use MuJoCo 🙄.
I was wondering if there is any work being done on adding the robotics environments?
If the answer to the above is no, I am willing to put in some time to get support for these environments. The only problem is, I am not 100% sure where to start, so any pointers would be nice.
Hey there! Unfortunately, the answer is no. But I can give you pointers on how to add a new environment within the next days. I should have added that information a long time ago anyways ;-)
Ok, so since I lacked the time until now to do a proper write-up, here is a quick write-up: It is basically as simple/hard as the following: Copy an env from the roboschool code and change the number of observations/actions to the number in the mujoco implementations of OpenAI. Then, and this is the most problematic part, find the observations from bullet which correspond to the ones in the mujoco implementations. Finally, to ensure compatibility, check if an agent trained in mujoco behaves similarly on your implementation. The last step could theoretically be a separate step, I would be fine if you found the corresponding observation signals.
I am also interested in moving fetch over also. I currently can't run RL sims on my university job based server because of Mujoco needing to be supported (which in the state of my university's robotics dept, that is never going to happen). They have a physical fetch robot, and I already want to edit the environments due to some requirements needed from my thesis, so plan to try my hand at assisting in the port.
Any update on this development?