Pictura · Jul 28, 2026
Pictura: Perspective-View Self-Play at Scale for Driving
Researchers introduced Pictura, a GPU-accelerated multi-agent driving simulator that renders each agent's egocentric view at every step, sustaining up to 500K agent-steps/s (2M images/s) on a single H100. Using Pictura, they trained Alberti, the first large-scale driving self-play policy trained directly from perspective images without privileged observations, spanning 50B agent steps (~35M km of driving). Alberti approaches the driving performance of its privileged vectorized counterpart.
What happened
A new simulator, Pictura, enables perspective-view self-play for autonomous driving at scale, rendering egocentric camera views for each agent to close the representation gap between privileged vectorized observations and real-world camera inputs. The resulting policy, Alberti, trained with PPO over 50 billion agent steps, demonstrates that self-play from raw images can approach the performance of policies relying on perfect state information.
Technical significance
Pictura achieves 500K agent-steps/s (2M images/s) on a single H100, enabling large-scale self-play directly from perspective images. Alberti is trained with plain PPO, showing that perspective-view self-play can mitigate the representation gap without distillation from privileged policies.
Industry impact
This work suggests that scalable, image-based self-play simulators could reduce reliance on expensive real-world data collection and privileged sensor setups for autonomous driving development.
What to watch
Next signals include open-sourcing of Pictura or Alberti, benchmarks against real-world driving datasets, and integration with commercial simulators. Further scaling of image-based self-play may lead to fully camera-only driving policies competitive with privileged systems.
Decision value
By enabling training of robust driving policies directly from camera inputs in simulation, Pictura could lower development costs and accelerate deployment of autonomous vehicles, reducing the need for costly real-world testing and specialized sensor suites.