taldatech/sketchy_128

Sketchy 128x128

This dataset is a pre-processed version of the Sketchy dataset (originally from DeepMind and the Open-X Embodiment effort) used in the paper Latent Particle World Models: Self-supervised Object-centric Stochastic Dynamics Modeling.

The data has been pre-processed to 128x128 pixels. Unlike the original TFRecords, this version consists of raw frames/images, facilitating use in standard PyTorch/JAX pipelines for training object-centric world models.

• Project Page: https://taldatech.github.io/lpwm-web
• GitHub Repository: https://github.com/taldatech/lpwm

Usage

To train the Latent Particle World Model (LPWM) on the Sketchy dataset, you can use the following command from the official repository:

python train_lpwm.py --dataset sketchy

Citation

@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}
}