Paper Overview
Research Area: Computer Vision Authors: Yuan Yin, Elias Ramzi, Marc Lafon, Valentin Charraut, Victor Bares, Yihong Xu, Éloi Zablocki, Alexandre Boulch, Thibault Buhet, Andrei Bursuc, Matthieu Cord Published: 2026-07-28 arXiv: 2607.26005
Summary
Self-play in simulation produces robust driving policies at scale. Previous demonstrations of such behavior have used privileged vectorized observations—such as exact poses and velocities, even for occluded agents. This approach assumes that perception is solved and introduces a representation gap with the partial observation of a deployed agent driving from the perspective view of egocentric cameras. A common fix—distilling the privileged policy into a camera-input student—leaves the student imitating decisions its own view cannot justify.
Instead, the authors establish perspective-view self-play as a practical training regime. They introduce Pictura, a GPU-accelerated multi-agent driving simulator that renders each agent's egocentric view at every step, mitigating the representation gap at its source.
Key Results
- Throughput: Pictura sustains up to 500K agent steps per second (2M images per second) on a single NVIDIA H100 GPU.
- Alberti: Trained with vanilla PPO self-play using Pictura, Alberti is the first large-scale driving self-play policy trained directly from perspective images, without privileged observations.
- Scale: Training spans 50 billion agent steps—approximately 35 million kilometers of driving.
- Performance: Alberti approaches the driving performance of its privileged vectorized counterpart.
- Zero-shot transfer: It transfers zero-shot to Waymo Open Motion Dataset layouts re-rendered in Pictura, where it outperforms privileged vectorized agents.
Original Abstract (excerpt)
> Self-play in simulation produces robust driving policies at scale. Demonstrations of such behavior have been made using privileged vectorized observations such as exact poses and velocities, even for occluded agents. This assumes that perception is solved and introduces a representation gap with the partial observation of a deployed agent driving from the perspective view of egocentric cameras. A common fix, distilling the privileged policy into a camera-input student, leaves the student imitating decisions its own view cannot justify. Instead, we establish perspective-view self-play as a practical training regime. We introduce Pictura, a GPU-accelerated multi-agent driving simulator that renders each agent's egocentric view at every step, mitigating the representation gap at its source.
*Auto-collected on 2026-07-30*