Unable to find KUKA dataset and pretrained models for training/evaluation

[shubhamsonkar01]

Hello,

Thank you for releasing the SafeDiffuser code.

I am currently working with the kuka branch and have successfully set up the environment. However, I am unable to locate the required KUKA dataset and pretrained models.

What I observed

1. The training script:

python scripts/kuka.py

fails inside KukaDataset because it expects:

dataset = "kuka_dataset/*.npy"

but the repository does not contain a kuka_dataset directory or any .npy trajectory files.

2. I searched the repository and found references to:

dataset = "/data/vision/billf/scratch/yilundu/pddlstream/output_5/*.npy"

which appears to be a local path from the original development environment.

3. I also ran:

./scripts/download_pretrained.sh

after installing gdown.

The download completed successfully (~2.25 GB), but the extracted contents only contain pretrained locomotion models for:

• hopper
• walker2d
• halfcheetah

I could not find:

• Maze2d pretrained models
• KUKA pretrained models
• kuka_dataset/*.npy
• any KUKA-specific checkpoints

Questions

1. Where can I obtain the KUKA expert trajectory dataset (kuka_dataset/*.npy) required for training?

2. Are the KUKAand Maze2d pretrained models publicly available? If yes, could you please provide the download link or instructions?

3. The README mentions that the training code automatically downloads the data on first use. Is there an additional script or branch required for this functionality?

Any guidance would be greatly appreciated.

[Weixy21]

All the pretrained models are provided by the diffuser paper (please check the diffuser repo). If some are not available, you can train by yourself.

You can track and locate the training data when you train the model. There is no additional script for downloading the data. It will be automatically downloaded for the first time training.