MimicLabs Study Datasets#
Download datasets#
We open-source all the datasets used in our study paper, which contains over 850k trajectories across 8 different scenes (lab1 through lab8) that were used to study the effects of different retrieval strategies from a large-scale simulation dataset. We have made low-dimensional observations and actions available for download. You can add image observations to these datasets using instructions in the subsequent sub-section.
Using web URL#
You can download all datasets from Hugging Face.
Using provided scripts (recommended)#
You can also download all datasets using the provided script (CAUTION: this will download ~115GB of data):
$ cd <PATH_TO_THIS_REPO>/mimiclabs/mimiclabs
$ python scripts/download_mimiclabs_datasets.py --download_dir <YOUR_DOWNLOAD_DIR>
By default the above script will download all datasets to the the download directory. You can selectively download datasets from a few labs. Below is an example command to download datasets of just lab1 and lab2:
$ python scripts/download_mimiclabs_datasets.py --download_dir <YOUR_DOWNLOAD_DIR> --labs lab1 lab2
If you’d like to do a dry run first, run:
$ python scripts/download_mimiclabs_datasets.py --download_dir <YOUR_DOWNLOAD_DIR> --labs lab1 lab2 --dry_run
Add image observations to datasets#
Downloaded datasets only contain low-dim observations. To add image observations to the downloaded datasets, run:
$ cd <PATH_TO_THIS_REPO>/mimiclabs/mimiclabs
$ python scripts/add_obs_to_mimiclabs_datasets.py --input_root_dir <YOUR_DOWNLOAD_DIR>/mimiclabs_study --output_root_dir <YOUR_DOWNLOAD_DIR>/mimiclabs_study
This will generate a script called add_obs_to_mimiclabs_datasets.sh in the scripts directory, which you can run using:
$ chmod +x scripts/add_obs_to_mimiclabs_datasets.sh
$ ./scripts/add_obs_to_mimiclabs_datasets.sh