ml-jku/meta-world

This repository contains the Meta-World datasets as used in Learning to Modulate Pre-trained Models in RL:

• The 2M folder contains trajectories (2M transitions) for every task are stored as separate .npz.
• The 2M_separate folder contains the same data, but every trajectory is stored as a separate .hdf5 file and every task is stored as a .tar.gz file.

Download the dataset using the huggingface-cli:

huggingface-cli download ml-jku/meta-world --local-dir=./meta-world --repo-type dataset

For dataloading we refer to our Github repository: https://github.com/ml-jku/L2M

Citation:

@article{schmied2024learning,
  title={Learning to Modulate pre-trained Models in RL},
  author={Schmied, Thomas and Hofmarcher, Markus and Paischer, Fabian and Pascanu, Razvan and Hochreiter, Sepp},
  journal={Advances in Neural Information Processing Systems},
  volume={36},
  year={2024}
}