This is a dataset of MQA task, you can get the simulation scene in V-REP, 3D object models, complicated scenes and corresponding question-anwer pairs.
- Paper MQA: https://arxiv.org/abs/2003.04641
- Source code:https://github.com/dengyh16code/MQA_ICRA2021
First, you need to unzip data.zip, then you will get:
├── README.md help
├── environment.py python file for vrep remote api
├── remoteApi.dll vrep configure file(Windows)
├── remoteApi.so vrep configure file(Linux)
├── scene.ttt verp simlation file
├── test.py demo to show a simulation scene of MQA task
├── vrep.py python file for vrep configure
├── vrepConst.py python file for vrep configure
├── data
│ ├── encode_ques encoding question file
│ ├── mesh 3D object models can be used in Vrep
│ ├── ques question file
│ ├── test_cases scenes file
│ ├── boundary_size.json size of 3D object\
│ ├── box.txt working space of UR5 in simulation
│ └── vocab.json vocabulary
We bulid our dataset in a simulation environment named V-REP. First, you need to open the our simulation scene file Simulation.ttt in V-REP.
Then you can use the python scipt test.py to load different scene in our dataset, where group_num can be taken from 0 to 9 and scene_num can be taken from 0 to 9. Different group_num and scene_num represent loading different scenes.
python test.py -group_num 1 -scene_num 0 
More remote api function(such as getting camera data, controlling UR5 manipulator, load questions) can be found in enviroment.py.
If you feel it useful, please cite:
@article{deng2020mqa,
title={MQA: Answering the Question via Robotic Manipulation},
author={Yuhong, Deng, and Xiaofeng, Guo, and Naifu, Zhang and Di, Guo and Huaping, Liu, and Fuchun, Sun},
journal={arXiv preprint arXiv:2003.04641},
year={2020}
}MIT License.