Haoran Zhu, Wancong Zhang, Yann LeCun, Anna Choromanska
New York University, AMI Labs
Timeline:
- Paper released on 09/28/2026
- Main source code released by 10/19/2026
- Collect and upload model weights to Hugging Face by 10/23/2026 (A bit busy these days, so pushing this back by 3 days. Sorry for the delay!)
- Review community feedback, fix bugs if needed, and refactor the codebase by 11/19/2026
The provided training and evaluation launchers are configured for GPU clusters managed by Slurm. The .sbatch scripts specify resource requests, environment setup, and filesystem paths. Adjust these settings to match your environment.
For standalone GPU servers, adapt the commands in these scripts into .sh launchers. We are testing standalone-server configurations and plan to provide updated instructions in the next release.
See the Installation Guide.
See the Training Guide.
See the Goal-Conditioned Zero-Shot Planning Evaluation Guide.
See the Downstream Imitation Learning Guide.
See the Downstream Imitation Learning Evaluation Guide.
@article{zhu2026ad,
title={AD-E2E-JEPA: A Joint-Embedding Predictive Architecture For End-to-End Autonomous Driving},
author={Zhu, Haoran and Zhang, Wancong and LeCun, Yann and Choromanska, Anna},
journal={arXiv preprint arXiv:2609.34085},
year={2026}
}