Pictura: Perspective-View Self-Play at Scale for Driving
The paper claims large-scale driving self-play can be trained directly from egocentric camera views, without privileged state inputs.
Pictura is a GPU-accelerated multi-agent simulator that renders each agent’s perspective view at every step. The authors report up to 500K agent-steps per second, or 2M images per second, on one H100. Using it, they train Alberti with PPO over 50B agent steps, about 35M km of simulated driving. The policy approaches a privileged vectorized counterpart and outperforms privileged agents in zero-shot transfer to Waymo layouts re-rendered in Pictura. ArXiv · AI/CL/LG's note
Pictura is a GPU-accelerated multi-agent simulator that renders each agent’s perspective view at every step. The authors report up to 500K agent-steps per second, or 2M images per second, on one H100. Using it, they train Alberti with PPO over 50B agent steps, about 35M km of simulated driving. The policy approaches a privileged vectorized counterpart and outperforms privileged agents in zero-shot transfer to Waymo layouts re-rendered in Pictura. ArXiv · AI/CL/LG's note
score 5