Datasets:

ArXiv:
License:
File size: 3,105 Bytes
7ae7f1b
 
 
8b5aad1
 
 
 
 
139eba7
 
9bcacb0
 
 
 
 
 
 
1701868
139eba7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8b5aad1
 
 
1701868
8b5aad1
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
---
license: cc-by-nc-sa-4.0
---

# Dataset Card for MimicGen Datasets

## Dataset Summary

This repository contains the official release of datasets for the [CoRL 2023](https://www.corl2023.org/) paper "MimicGen: A Data Generation System for Scalable Robot Learning using Human Demonstrations".

The datasets contain over 48,000 task demonstrations across 12 tasks, grouped into the following categories:
- **source**: 120 human demonstrations across 12 tasks used to automatically generate the other datasets
- **core**: 26,000 task demonstrations across 12 tasks (26 task variants)
- **object**: 2000 task demonstrations on the Mug Cleanup task with different mugs
- **robot**: 16,000 task demonstrations across 4 different robot arms on 2 tasks (4 task variants)
- **large_interpolation**: 6000 task demonstrations across 6 tasks that pose significant challenges for modern imitation learning methods

For more information please see the [website](https://mimicgen.github.io), the [paper](https://arxiv.org/abs/2310.17596), and the [code](https://github.com/NVlabs/mimicgen_environments).

## Dataset Structure

Each dataset is an hdf5 file that is readily compatible with [robomimic](https://robomimic.github.io/) --- the structure is explained [here](https://robomimic.github.io/docs/datasets/overview.html#dataset-structure). 

As described in the paper, each task has a default reset distribution (D_0). Source human demonstrations (usually 10 demos) were collected on this distribution and MimicGen was subsequently used to generate large datasets (usually 1000 demos) across different task reset distributions (e.g. D_0, D_1, D_2), objects, and robots.

The datasets are split into different types:

- **source**: source human datasets used to generate all data -- this generally consists of 10 human demonstrations collected on the D_0 variant for each task.
- **core**: datasets generated with MimicGen for different task reset distributions. These correspond to the core set of results in Figure 4 of the paper.
- **object**: datasets generated with MimicGen for different objects. These correspond to the results in Appendix G of the paper.
- **robot**: datasets generated with MimicGen for different robots. These correspond to the results in Appendix F of the paper.
- **large_interpolation**: datasets generated with MimicGen using much larger interpolation segments. These correspond to the results in Appendix H in the paper.

**Note**: We found that the large_interpolation datasets pose a significant challenge for imitation learning, and have substantial room for improvement.

## Citation

Please cite the [MimicGen paper](https://arxiv.org/abs/2310.17596) if you use these datasets in your work:

```bibtex
@inproceedings{mandlekar2023mimicgen,
    title={MimicGen: A Data Generation System for Scalable Robot Learning using Human Demonstrations},
    author={Mandlekar, Ajay and Nasiriany, Soroush and Wen, Bowen and Akinola, Iretiayo and Narang, Yashraj and Fan, Linxi and Zhu, Yuke and Fox, Dieter},
    booktitle={7th Annual Conference on Robot Learning},
    year={2023}
}
```