Parameters used to replicate training the demo checkpoint?

[fionalluo]

Hi, I was wondering what settings were used to obtain the example pretrained checkpoint?

I ran the example commands in the README

for cube_scale in 0.9 0.95 1.0 1.05 1.1 
do
	bash scripts/gen_grasp.sh $cube_scale custom_grasp_cache num_envs=1024 
done

python3 train.py task=LeapHandRot max_iterations=1000 task.env.grasp_cache_name=custom_grasp_cache

However, the checkpoint I trained from these commands looks significantly worse than the example checkpoint. (I also rendered the example checkpoint and verified that it matches the provided video.) Were the training settings used to obtain the example policy different, or is it a different issue with my setup? I couldn't find them in the paper. Here's a video of the policy I trained with the example commands. Thank you!

policy_vid.mov