---
license: cc-by-4.0
task_categories:
- robotics
tags:
- robot
- ogbench
- rl
- imitation
- learning
- simulation
- manipulation
---

# OGBench Data for Latent Particle World Models (LPWM)

This repository contains pre-processed 64x64 frames for the `scene` and `cube` tasks from the [OGBench benchmark](https://github.com/seohongpark/ogbench). The dataset includes actions and frames used for training and evaluating **Latent Particle World Models (LPWM)**.

LPWM is a self-supervised object-centric world model that autonomously discovers keypoints, bounding boxes, and object masks directly from video data. It is designed to scale to real-world multi-object datasets and is applicable in decision-making tasks such as goal-conditioned imitation learning.

- **Paper:** [Latent Particle World Models: Self-supervised Object-centric Stochastic Dynamics Modeling](https://huggingface.co/papers/2603.04553)
- **Project Page:** [https://taldatech.github.io/lpwm-web](https://taldatech.github.io/lpwm-web)
- **GitHub Repository:** [https://github.com/taldatech/lpwm](https://github.com/taldatech/lpwm)

## Citation

If you use this data or the LPWM model in your research, please cite the following paper:

```bibtex
@inproceedings{
  daniel2026latent,
  title={Latent Particle World Models: Self-supervised Object-centric Stochastic Dynamics Modeling},
  author={Tal Daniel and Carl Qi and Dan Haramati and Amir Zadeh and Chuan Li and Aviv Tamar and Deepak Pathak and David Held},
  booktitle={The Fourteenth International Conference on Learning Representations},
  year={2026},
  url={https://openreview.net/forum?id=lTaPtGiUUc}
}
```