task_categories:
- robotics
tags:
- code
size_categories:
- 100B<n<1T
Robotic Manipulation Datasets for Four Tasks
[Project Page] [Paper] [Code] [Models] [Raw GoPro Videos]
This repository contains in-the-wild robotic manipulation datasets collected using UMI, and processed through a SLAM pipeline, as described in the paper "Data Scaling Laws in Imitation Learning for Robotic Manipulation". The datasets cover four tasks:
- Pour Water
- Arrange Mouse
- Fold Towel
- Unplug Charger
Dataset Folders:
arrange_mouse and pour_water: Each folder contains data from 32 unique environment-object pairs, with 120 demonstrations per pair.
fold_towel and unplug_charger: Each folder contains data from 32 unique environment-object pairs, with 60 demonstrations per pair.
pour_water_16_env_4_object and arrange_mouse_16_env_4_object: These folders contain data from 16 environments, with 4 different manipulation objects per environment, and 120 demonstrations per object.
Note that due to the size of the pour_water_16_env_4_object/dataset.zarr.zip file (over 50GB), it has been split into two parts. You can restore the full dataset using the following command:
cat pour_water_16_env_4_object/dataset_part_* > pour_water_16_env_4_object/dataset.zarr.zip
Additional Information
- Each dataset is a merge of smaller datasets (one per environment-object pair). Inside each folder, you will find a count.txt file that lists the number of demonstrations in each smaller dataset.
- These datasets can be used to train policies that generalize effectively to novel environments and objects.
- For more details on how to use our datasets, please refer to our code.