Xuanlin (Simon) Li

Was able to reproduce it for point cloud BC. Though for RGB-D BC, the gradient does not fall to zero. (RGB-D BC also requires more memory).

Hi Erik, Thanks for your detailed checking of our code! The target critic is intentionally designed to share the backbone; otherwise the training is unstable.

Could you try interactive python: `CUDA_VISIBLE_DEVICES=0 python` and run the following: ``` import mani_skill2.envs, gym env=gym.make('Excavate-v0', obs_mode='pointcloud') obs=env.reset() ```

Could you reinstall maniskill2 from scratch in a new conda environment? This should be maniskill2 install problem instead of maniskill2-learn. Also did you `pip install warp`? The warp referred by...

Hi MDMLab223, Has the issue been resolved yet?

ViT is known to suffer significant overfitting under a small training dataset even under strong augmentations, as it does not utilize priors such as translation equivariance in CNN. For it...

I had a similar issue and I solved it by replacing the problematic functions that use `torch.repeat_interleave` in the original SAM repo with monkey patch, see #78