Pictura_arxiv

New preprint: Pictura is a GPU-accelerated multi-agent driving simulator that renders every agent’s egocentric view at each step, sustaining up to 500K agent-steps/s on a single H100. With it, we train Alberti by self-play with plain PPO over 50B agent steps (~35M km): the first large-scale driving self-play policy learned directly from perspective images, with no privileged observation of the surroundings.