Apollo World

Apollo World is an agent-driven, omni-modal world model for physical AI first unveiled by Baidu Apollo at Baidu AI Day · Apollo in Shenzhen on 2026-11-03, and widely covered by Xinhua, Yicai and InfoQ on Nov 4–5.

Rather than a pretty video generator, it acts as an examiner for driverless cars: in its virtual cities, cars, pedestrians and cyclists are AI agents that make decisions and react to each other. It generates camera frames, LiDAR point clouds and other sensor data in real time so the driving stack can perceive, decide and act as on real roads, completing full order → pickup → drop-off trips in closed loop. Baidu says it is already used for global testing and validation of Apollo Go (Luobo Kuaipao) and can simulate about 100,000 km per day.

What problem it solves

Autonomous driving has to balance safety, intelligence and scale at once: long-tail scenarios are hard to cover with real miles, entering a new city is costly and risky, and continuous cross-segment driving needs large-scale validation. Apollo World moves much of that testing into a virtual world first.

The "2+3" criteria

• 2 baseline requirements:
  • Causality: predict the future from the past, not stitch frames together;
  • Interactivity: respond to injected actions and change future world state.
• 3 advanced metrics: physical accuracy, persistence (no drift over long runs) and real-time performance (low latency).

Three core capabilities

1. Agent-driven: traffic participants are not scripted props; they react differently when the robotaxi slows, changes lanes or yields.
2. Omni-modal perception: real-time camera and LiDAR observations that change consistently with weather.
3. Trip-level closed loop: full order-level simulation between arbitrary start and end points, not isolated intersection clips.

Limitations

• Currently an internal Baidu stack for Apollo Go; no public API or weights.
• Baidu says it can extend to robotics and other physical-AI settings, but there is no public product yet.

Who it's for

• Researchers and engineers working on driving simulation, world models and physical AI;
• Investors and analysts tracking robotaxi technology roadmaps.

How to follow it

1. Read Xinhua's report (Baidu unveils Apollo World) and InfoQ's technical write-up.
2. Compare it with other world models: Apollo World emphasizes interactivity and closed-loop validation, not video quality.
3. Watch for a public simulation platform, paper, or robotics extension from Baidu Apollo.