---
title: "Pictura: Perspective-View Self-Play at Scale for Driving"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2607.26005v1"
permalink: "https://feed.astralor.com/items/itm_cbca33f2865600f74eb98f74"
published: "2026-07-28T17:20:39.000Z"
score: 81
channel: "ai"
tags: ["Alberti","Pictura","PPO","Autonomous Driving","Self-Play","Perspective View"]
---

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

## AI commentary

> This paper truly achieves 'what you see is what you get'—letting agents learn directly from their own camera view instead of relying on privileged distillation. For researchers in autonomous driving simulation and reinforcement learning, Pictura's high throughput and Alberti's zero-shot transfer capability are highly instructive. Notably, it demonstrates that pure vision-based self-play can match or even surpass privileged methods at reasonable computational cost, which may reshape future driving policy training paradigms.

## At a glance

Pictura is a GPU-accelerated multi-agent driving simulator that sustains up to 500K agent-steps/s (2M images/s) on a single H100. Using this simulator, researchers trained Alberti, a driving policy, via plain PPO self-play directly from perspective images without any privileged observations (e.g., exact poses or velocities). The training spanned 50B agent steps (~35M km of driving). Alberti approaches the performance of privileged vectorized counterparts and zero-shot transfers to Waymo Open Motion Dataset layouts re-rendered in Pictura, even outperforming privileged vectorized agents. This work bridges the representation gap between simulation training and real-world deployment.

## Key points

- Pictura achieves 500K agent-steps/s on a single H100, rendering a perspective view for each agent at every step.
- Alberti policy is trained via self-play with plain PPO using only perspective images, totaling 50B steps (~35M km).
- Alberti zero-shot transfers to Waymo Open Motion Dataset layouts, outperforming privileged vectorized agents.
- This approach learns directly from sensor input, eliminating the representation gap caused by privileged distillation.

## Full text

Self-play in simulation produces robust driving policies at scale. Demonstrations of such behavior have been made using privileged vectorized observations such as exact poses and velocities, even for occluded agents. This assumes that perception is solved and introduces a representation gap with the partial observation of a deployed agent driving from the perspective view of egocentric cameras. A common fix, distilling the privileged policy into a camera-input student, leaves the student imitating decisions its own view cannot justify. Instead, we establish perspective-view self-play as a practical training regime. We introduce Pictura, a GPU-accelerated multi-agent driving simulator that renders each agent's egocentric view at every step, mitigating the representation gap at its source. Pictura sustains up to 500K agent-steps/s (2M images/s) on a single H100. Using Pictura, we train Alberti by self-play with plain PPO. It is the first large-scale driving self-play policy trained directly from perspective images, without privileged observations. Training spans 50B agent steps for ~35M km of driving. It approaches the driving performance of its privileged vectorized counterpart, and transfers zero-shot to Waymo Open Motion Dataset layouts re-rendered in Pictura, where it outperforms privileged vectorized agents. Project page: https://valeoai.github.io/Pictura/

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_cbca33f2865600f74eb98f74