metadata
license: cc-by-4.0
task_categories:
- robotics
tags:
- simulation
- panda
- isaac-gym
- robot
- manipulation
- push-t
- imitation-learning
- world-models
Panda Dataset
Project Page | Paper | GitHub
This repository contains expert imitation trajectories (actions and frames) for several robotic manipulation tasks, as presented in the paper Latent Particle World Models: Self-supervised Object-centric Stochastic Dynamics Modeling.
Dataset Description
The data consists of expert trajectories generated using a Panda robot in the IsaacGym simulator. It was used to evaluate Latent Particle World Models (LPWM) in tasks requiring stochastic dynamics modeling and goal-conditioned imitation learning.
- Tasks:
- Cube: Manipulation of 1, 2, or 3 cubes.
- Push-T: Pushing T-shaped objects (1, 2, or 3 T's).
- Observations: 128x128 resolution RGB frames, including 2 camera views per task.
- Actions: Continuous robot actions corresponding to the trajectories.
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}
}