Physix FrontierPhysix Frontier · News Briefing Card (Leiphone · Nov 5, 2026)
Baidu unveils ApolloWorld world model with 2+3 evaluation criteria
KEY FACTS
- Baidu disclosed its world model ApolloWorld, which models traffic participants as AI agents that can respond to driving actions.
- ApolloWorld can generate camera images and LiDAR point clouds simultaneously, and control scenes by conditions such as weather and lighting.
- Baidu proposed the "2+3" standard for world models: causality and interactivity, plus physical accuracy, persistence, and real-time performance.
- Apollo Go has accumulated over 350 million kilometers of autonomous driving mileage, of which over 240 million kilometers were fully driverless.
- Baidu says one day of virtual mileage in ApolloWorld is roughly equivalent to three consecutive years of real-vehicle testing with a safety driver.
KEY DATA
350 million kmApollo Go cumulative autonomous driving mileage
240 million kmApollo Go fully driverless mileage
over 1 million kmBaidu end-to-end model fully driverless operating mileage
three yearsVirtual one day equivalent real-vehicle test duration
PHYSIX OBSERVATION
The competitive focus of world models is shifting from visual realism to causal trustworthiness. By putting evaluation criteria on the table, Baidu aims to seize the narrative in simulation. For the industry, cloud-based parallel testing can lower the real-vehicle cost of expanding into new cities; but demos are mostly success cases, and coverage of failure scenarios is the real yardstick for this testing ground.
Source: Leiphone report