Skip to content

README.md

Latest commit

 

History

History
61 lines (44 loc) · 1.64 KB

README.md

File metadata and controls

61 lines (44 loc) · 1.64 KB

Dataset

medium-replay dataset

  • MetaWorld medium-replay dataset

  • DMControl medium-replay dataset

We used PEBBLE repository to create new offline preference-based RL dataset. We collected the replay buffer from SAC training using GT rewards implemented in PEBBLE.

You can download dataset here (Google Drive)

or automatically download:

bash ./download.sh

dataset directory should look like:

dataset
├── MetaWorld
│  ├── box-close-v2
|  │  ├── saved_replay_buffer_1000000_seed0
|  │  ├── saved_replay_buffer_1000000_seed1
|  │  ├── saved_replay_buffer_1000000_seed2
│  └── button-press-topdown-v2
│  └── dial-turn-v2
│  └── ...
├── DMControl
│  ├── cheetah-run
│  └── hopper-hop
│  └── humanoid-walk
│  └── ...

medium-expert dataset

For metaworld medium-expert dataset, you can use IPL repository

Human feedback dataset

You can download human_feedback/metaworld_button-press-topdown-v2/dataset.pkl dataset here (Google Drive)

human_feedback directory should look like:

human_feedback
├── metaworld_button-press-topdown-v2
│  ├── dataset.pkl
│  ├── _Independent.txt
│  ├── _SeqRank.txt
│  ├── _RLT_0.txt
│  ├── _RLT_1.txt
│  ├── render_xxx.gif
│  ├── ...
...