rubiks-dataset

Dataset to assess the differences between the unsupervised learning methods ICM and M²VAE

If you use this dataset in your work, please cite it as follows:

Bibtex

@misc{korthals2019perceived,
archivePrefix = {arXiv},
arxivId = {cs.RO/1911.00584},
author = {Korthals, Timo and Schilling, Malte and Leitner, Jürgen},
eprint = {1911.00584},
pages = {2},
primaryClass = {cs.RO},
title = {{A Perceived Environment Design using a Multi-Modal Variational Autoencoder for learning Active-Sensing}},
url = {https://sites.google.com/site/dpgmcar2019/home},
year = {2019}
}

Description

Check out the YouTube video to see, how the data was generated.

rubiks is a dataset of three different views on a Rubik cube which where sampled by performing viewpoint changes with a Franka Emica Panda robot.

Latent factor values

• Viewpoint (impact on the background): ['left', 'front', 'right']
• Rubik's state (impact on the foreground): [['blue', 'red', 'green'], ...]
• Actions (from -> to): [['left', 'front'], ['left','right'], ...]

All possible Rubik states for the corresponding viewpoints

Rubik's state	left	front	right
1	blue	red	green
2	blue	yellow	green
3	blue	orange	green
4	blue	white	green
5	green	red	blue
6	green	yellow	blue
7	green	orange	blue
8	green	white	blue
9	white	orange	yellow
10	white	red	yellow
11	white	green	yellow
12	white	blue	yellow
13	yellow	red	white
14	yellow	orange	white
15	yellow	green	white
16	yellow	blue	white
17	orange	green	red
18	orange	blue	red
19	orange	yellow	red
20	orange	white	red
21	red	green	orange
22	red	blue	orange
23	red	yellow	orange
24	red	white	orange

All actions and corresponding action representation

Action	Viewpoint 1	Viewpoint 2	left-state	front-state	right-state
1	left	front	0	1	-1
2	left	right	0	-1	1
3	front	left	1	0	-1
4	front	right	-1	0	1
5	right	left	1	-1	0
6	right	front	-1	1	0