zavierand/CAPs

Contact-Anchored Policies: Contact Conditioning Creates Strong Robot Utility Models - Dataset

Zichen Jeff Cui¹, Omar Rayyan³, Haritheja Etukuru², Bowen Tan¹, Zavier Andrianarivo¹, Zicheng Teng¹, Yihang Zhou¹, Krish Mehta⁶, Nicholas Wojno¹, Kevin Yuanbo Wu¹, Manan H. Anjaria¹, Ziyuan Wu¹, Manrong Mao¹, Guangxun Zhang¹, Binit Shah⁴, Yejin Kim⁵, Soumith Chintala¹, Lerrel Pinto¹, and Nur Muhammad Mahi Shafiullah²

¹ New York University, ² UC Berkeley, ³ UCLA, ⁴ Hello Robot Inc., ⁵ Ai2, ⁶ University of Waterloo

This is the official CAPs dataset used to train CAPs policies. It is released under the MIT license.

Overview of Dataset

The root of the dataset contains the raw dataset used for training as well as a soon-to-be-released LeRobot format of the dataset. The raw dataset contains the three tasks trained in the paper — Pick, Open, and Close.

Each task contains the following trajectories:

	Pick	Open	Close
Trajs	14,606	7,380	4,152

In total, there are 26,138 trajectories totaling ~23 hours of human-collected data.

Dataset Structure

Each task (Pick, Open, Close) follows the same layout:

<Task>/
└── home_xx/                  # anonymized home ID (may include _flipped)
    └── <env>/                # environment (e.g. Env1, lemon-pickup-x169)
        └── <timestamp>/      # demonstration timestamp
            ├── compressed_video_h264.mp4
            └── labels.json

And at the dataset root:

dataset_raw/
├── Pick/
├── Open/
└── Close/

Collecting Your Own Data: To collect your own data, the STL files for our handheld-collection grippers have also been open-sourced and released in an effort to further robot-learning research. Collecting your own data is as simple as collecting with the AnySense app and storing your data in the format structure provided above.

Download the Dataset

To download the dataset, you can use GitHub's clone command to download the dataset into a directory of your choice:

# clone entire repo
mkdir path/to/store/dataset && cd path/to/store/dataset
git clone https://huggingface.co/datasets/zavierand/CAPs

This will clone from the root and should take ~ 40 minutes to clone.

Installation (Raw Dataset)

1. Get Conda/Mamba Running

From the CAPs docs, clone the repo, cd into the project directory, and install the conda environment home_robot.

git clone https://github.com/jeffacce/cap-policy
cd cap-policy

conda env create -f conda_env.yaml

conda activate home_robot

# optionally, mamba is supported as well
# mamba env create -f conda_env.yaml
# mamba activate home_robot

2. Set Environment Variables

From the project root, run setup.sh to configure environment variables:

# in cap-policy

./setup.sh

3. Set Dataset Paths

After installing the home_robot environment and setting up your environment variables, set the data_root.train and data_root.val variables to the root path of your dataset. The environment variables config file should be at the path cap-policy/imitation-in-homes/configs/env_vars/env_vars.yaml.

Your env_vars.yaml file should look something like this after the changes:

project_root: /path/to/folder/imitation-in-homes

data_root:
  - train: /path/to/dataset_raw/
  - val: /path/to/dataset_raw/

data_original_root:
  - train: /path/to/dataset_raw/
  - val: /path/to/dataset_raw/

Note: When cloning, if the whole repo is cloned and saved under CAPs and you wish to train on the raw dataset, you will need to set the environment variables in env_vars.yaml to the path of the dataset_raw directory.

Training

Now, you're ready to run a training job! Training documentation will be updated. Check the official GitHub repo for updates consistently on when training documentation will be up. An update will also be provided here once training documentation is up!

BibTeX

If you use our dataset, codebase, or models, please consider citing! Contributions are welcome!

@article{cui2026contact,
title={Contact-Anchored Policies: Contact Conditioning Creates Strong Robot Utility Models},
author={Zichen Jeff Cui and Omar Rayyan and Haritheja Etukuru and Bowen Tan and Zavier Andrianarivo and Zicheng Teng and Yihang Zhou and Krish Mehta and Nicholas Wojno and Kevin Yuanbo Wu and Manan H Anjaria and Ziyuan Wu and Manrong Mao and Guangxun Zhang and Binit Shah and Yejin Kim and Soumith Chintala and Lerrel Pinto and Nur Muhammad Mahi Shafiullah},
journal={arXiv preprint arXiv:2602.09017},
year={2026}
}