Inconsistency between the training data and the eval’s image rendering in MetaWorld.

[AdaChambers]

When I run python eval.py agent=baku suite=metaworld dataloader=metaworld suite/task=metaworld suite.hidden_dim=256 bc_weight=weights/metaworld/baku.pt use_proprio=false I find that the image of the eval rollout is like:

However the image in the metaworld expert_demo dataset is like:

The inconsistency may cause bad performance (almost 0% success rate) of the pretrained model. So a natural question is whether the Metaworld‘s env_wrapper rendering code is missing some image cropping operations or something else? Looking forward to anyone's response!

[ATriplez]

Hi, I had the same problem at first. At last I found out that the cause was my own installation of metaworld rather than using the version provided by the author which is below. I hope it will be helpful.

https://github.com/siddhanthaldar/Metaworld/tree/65a7c33c8326cadfef7e19590af98b2bf2640cb7

[siddhanthaldar]

I am closing this issue since it seems solved.