Pictura: GPU-Accelerated Driving Simulator for Perspective-View Self-Play
A recent study published on arXiv presents Pictura, a multi-agent driving simulator enhanced by GPU acceleration that generates each agent's egocentric perspective at every moment, facilitating self-play training for effective driving strategies. This method addresses the representation gap between detailed vectorized observations (precise poses and velocities) and limited camera-input observations. Pictura can achieve up to 500,000 agent-steps per second (equating to 2 million images per second) on a single NVIDIA H100 GPU. The authors utilize Pictura to train Alberti through self-play using standard PPO (Proximal Policy Optimization), marking the creation of the first driving policy developed solely from egocentric camera views without the use of privileged data or distillation.
Key facts
- Pictura is a GPU-accelerated multi-agent driving simulator.
- It renders each agent's egocentric view at every step.
- Pictura achieves up to 500K agent-steps/s (2M images/s) on a single H100.
- The paper proposes perspective-view self-play as a practical training regime.
- Alberti is trained using Pictura with plain PPO.
- The approach avoids the representation gap between privileged and camera observations.
- Previous methods used privileged vectorized observations like exact poses and velocities.
- The paper is available on arXiv with ID 2607.26005.
Entities
Institutions
- arXiv