Skip to content
Switch branches/tags

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

Home Grasping Dataset

Tools for downloading and using the home grasping dataset in the Robot Learning in Homes paper. In this pruned dataset, there are 28,874 grasp attempts with a 18.6% grasp success rate.

Download links

There are three datasets you can work with from this repository:

Data organization

The raw grasping data is organized as follows:

├── grasping_data
│   ├── <ROBOT_ID>
│   │   │   ├── grasp_<ATTEMPT_ID>
│   │   │   │   ├── color.jpg
│   │   │   │   ├── depth.jpg
│   │   │   │   ├── data.p
└── .   .   .   .

ROBOT_ID runs from 1 to 5 and corresponds to the robot on which the data was collected. GRASP_ENVIRONMENT corresponds to a grasping run. All the grasps in a specific environment run are in the same home. ATTEMPT_ID corresponds to a specific grasp attempt in the GRASP_ENVIRONMENT. Each grasp contains the RGB color.jpg image, the Depth depth.jpg image, and a pickle file data.p that contains information about the grasp attempt (like success or failure).

Getting started with the data

To get started with this data, we will use the smaller subsample. To get the full data, change the wget download to

cd ~
git clone
cd home_dataset
tar -xvzf grasping_data_small.tar.gz

To read the data, a few helper scripts are provided in scripts. To use them you will need to install appropriate python packages on python 2.7:

sudo apt-get install python-numpy python-scipy
pip install cloudpickle ipython==5 opencv-python temp argparse pytest-shutil

Run the following script to display and extract the patch dataset from the raw grasping data.

python scripts/ --home_dataset_path ~/home_dataset/grasping_data_small --patch_dataset_path '/tmp/' --train_fraction 0.8 --display 1 --msec 1000

To download pre-extracted patch dataset:

cd ~/home_dataset
tar -xvzf Patch_Dataset.tar.gz

To visualize the patch dataset, run:

python scripts/ --patch_dataset_path ~/home_dataset/patch_dataset/Train --pos 1 --rand 0 --msec 1000


The grasp data used is from the Robot Learning in Homes paper.

  title={Robot learning in homes: Improving generalization and reducing dataset bias},
  author={Gupta, Abhinav and Murali, Adithyavairavan and Gandhi, Dhiraj Prakashchand and Pinto, Lerrel},
  booktitle={Advances in Neural Information Processing Systems},


A grasping dataset collected in homes



No releases published


No packages published