Official Repository for Gigapixel

Scaling Self-Play for End-to-End Driving
Luke Rowe^1,2, Roger Girgis^1,3,4, Rodrigue de Schaetzen^1,2,4, Daphne Cornelisse⁵, Alaap Grandhi^1,6, Felix Heide^4,7, Eugene Vinitsky⁵, Christopher Pal^1,2,3, Liam Paull^1,2
1 Mila, ² Université de Montréal, ³ Polytechnique Montréal, ⁴ Torc Robotics, ⁵ NYU Tandon School of Engineering, ⁶ McMaster University, ⁷ Princeton University

arXiv preprint, 2026

Code will be released soon.

Gigapixel is a high-throughput batched driving simulator with perspective rendering that enables large-scale self-play directly from pixels. We use Gigapixel to train end-to-end driving policies via large-scale self-play directly from pixels; these policies transfer to real-world sensor data through lightweight perception adaptation, without human trajectory supervision.

world_0013.mp4

Table of Contents

• Citation

Citation

@article{rowe2026gigapixel,
  title   = {Scaling Self-Play for End-to-End Driving},
  author  = {Rowe, Luke and Girgis, Roger and de Schaetzen, Rodrigue and Cornelisse, Daphne and Grandhi, Alaap and Heide, Felix and Vinitsky, Eugene and Pal, Christopher and Paull, Liam},
  journal = {arXiv preprint arXiv:2606.19641},
  year    = {2026}
}