Training Mujoco-Gym Tasks with Image Observations

[nileshop22]

Hi @justinjfu @aviralkumar2907, I want to train Mujoco-Gym Continuous Control Tasks with Image Observations. I figured out that image_envs branch support Image Observations. But when I did gym.make('hopper-random-vision-v1') I got numerous issues. First one was NameError, Image wasn't defined in the img_utils.py file. Secondly, I got this error KeyError: ((1, 1, 3), '<f8') and I'm not sure why this error is coming. Do you have any running and well tested code to train Offline RL algorithms with D4RL with Images?

Also, have you ever tried training Mujoco-Gym tasks with Images? If yes, how did the different algorithms performed. If no, do you think they will perform good enough from just images?

Also, is this library compatible with new free Mujoco version 2.1.0?

Any help would be really appreciated!

Thanks!

[return-sleep]

Have you solved the problem yet?