Pictura: Perspective-View Self-Play at Scale for Driving

Researchers introduced Pictura, a GPU-accelerated multi-agent driving simulator that renders each agent’s first-person camera view at every step.

Using Pictura, they trained Alberti with standard PPO on 50 billion agent steps from egocentric views, without privileged state inputs.

Alberti reached near-parity with a privileged vector-based policy and improved zero-shot transfer on Waymo re-rendered scenarios.

The work suggests camera-based driving policies can be trained at scale, narrowing the gap between simulation and real-world deployment.