README.md

---
task_categories:
- robotics
tags:
- humanoid
- world-model
- contact-planning
- ego-vision
---

# EgoVCP Dataset

This repository contains the dataset for the paper [Ego-Vision World Model for Humanoid Contact Planning](https://huggingface.co/papers/2510.11682).

[**Project Page**](https://ego-vcp.github.io/) | [**GitHub**](https://github.com/HybridRobotics/Ego-VCP) | [**arXiv**](https://arxiv.org/abs/2510.11682)

The EgoVCP dataset is a demonstration-free offline dataset used to train a learned world model for humanoid contact planning. It features data collected from a G1 humanoid robot in the Isaac Lab simulation environment, enabling the robot to predict outcomes in a compressed latent space for contact-aware tasks.

## Dataset Tasks
The dataset includes rollouts for three primary manipulation and navigation scenarios:
- **Wall Support (`wall`)**: Supporting the robot against a wall after a perturbation.
- **Ball Blocking (`ball`)**: Interacting with and blocking incoming objects (e.g., a yoga ball).
- **Tunnel Traversal (`tunnel`)**: Traversing height-limited arches or tunnels.

## Dataset Structure
The dataset contains:
- **Ego-centric Depth Images**: Captured from the robot's perspective.
- **Proprioception Data**: Internal robot state information.
- **Actions and Rewards**: For offline world model training.

The data is organized by task into subdirectories: `wall`, `ball`, and `tunnel`.

## Usage
To use this dataset with the official implementation, you can clone it directly into your project directory:

```bash
cd Ego-VCP
mkdir dataset
git clone https://huggingface.co/datasets/Hang917/EgoVCP_Dataset.git dataset
```

## Citation
If you use this dataset in your research, please cite the following paper:
```bibtex
@article{liu2025egovcp,
  title={Ego-Vision World Model for Humanoid Contact Planning},
  author={Liu, Hang and others},
  journal={arXiv preprint arXiv:2510.11682},
  year={2025}
}
```