Baidu Apollo World unveiled: agent-based traffic flow and '2+3' criteria, no third-party access mentioned
According to a November 3, 2026 report by Leiphone, Baidu Apollo first disclosed its world model Apollo World at Shenzhen AI Day, saying it has been used in Apollo Go's global testing and proposing '2+3' evaluation criteria; the report did not mention a public API, weights or pricing.

What changed
According to a November 3, 2026 report by Leiphone, Baidu Apollo first disclosed Apollo World, a world model for physical AI, at the Baidu AI Day Apollo special session in Shenzhen. Baidu says the model is agent-driven, can generate omni-modal observations and supports a complete trip closed loop. Wang Liang, chief architect of Baidu IDG, said a world model should not only pursue visual realism but should serve as an 'examiner' for validating autonomous driving safety. Yang Ze, head of the world model, said Apollo World models traffic participants such as vehicles and pedestrians in virtual scenarios as AI agents with decision-making and interaction capabilities; it can generate observation data from cameras, LiDAR and other sensors in real time, and allow autonomous vehicles to continuously complete order acceptance, passenger pickup and drop-off tasks between different origins and destinations in a virtual city. Baidu also proposed '2+3' criteria: causality and interactivity as baseline thresholds, and physical accuracy, persistence and real-time performance as advanced indicators. The report says the model has been used in Apollo Go's global testing and validation; Apollo Go has operations in 28 cities, more than 23 million trips and over 350 million kilometers of autonomous driving. Real-road data is used to extract high-value scenarios, and virtual test results then feed back into vehicle-side iteration. The report did not disclose a public API, weights, license or pricing for Apollo World.
What it means for developers
For most API users, this news gives no accessible endpoint, SDK, weights or billing information, so it cannot be treated as a language-model substitute. For autonomous driving and robotics teams, if access becomes available in the future, the integration work is not a single model call but a combination of simulation environment, scenario orchestration, sensor data pipeline and safety evaluation process. This means existing pipelines that only replay real logs likely need to add agent behavior configuration, multimodal generation interfaces and long-task orchestration. If the reported agent-based traffic flow and omni-modal observations hold, the testing bottleneck will shift from image realism to interaction reproducibility, generation latency and long-run drift. For example, traditional scripted traffic flow suits regression testing because the same scenario can be repeated; agent-based traffic flow is better at discovering interactive long-tail cases but has higher variance, so fixed seeds, stratified sampling and confidence intervals must be used to judge whether improvements are real. Omni-modal generation means perception teams cannot look only at image quality; they must also check the geometric consistency of generated LiDAR, latency and the domain gap of downstream detectors. The reported '100,000 km per day' is an industry figure's analysis, not a reproducible benchmark; it cannot replace measurements of effective miles per unit time, scenario coverage and safety intervention rate. For robotics teams, the report only says the tech stack can be extended, without migration evidence, so it cannot be judged as directly reusable.
Integration and evaluation
The following are conditional recommendations based on the report, not tested conclusions about Apollo World. Decision one: if your team needs a stable regression baseline, keep scripted traffic flow as the main regression set and agent-based traffic flow as an exploration set; the trigger is needing to compare takeover rates across versions on the same route. Measurable metrics are the consistency rate of scenario outcomes under the same random seed, takeovers per thousand kilometers and the width of confidence intervals across multiple runs. Decision two: if you want to evaluate whether the 'complete trip closed loop' is worth integrating, first change the test unit from a single scenario to a multi-segment task including order acceptance, pickup, driving and drop-off; the precondition is that you can record full-trip sensor data and have human or remote safety operators as fallback. Measurable metrics are task completion rate, manual interventions per hundred trips, pickup and drop-off success rates and ride comfort indicators. Decision three: if your perception stack relies on generated data, require suppliers to separately report generation latency distributions for cameras and LiDAR, long-run drift, and detector gaps on real versus simulated data; only expand the share of generated data when the downstream gap from real road tests can be quantified. Decision four: treat '2+3' as a question checklist rather than a procurement conclusion, and ask the provider for operational definitions and test methods for causality, interactivity, physical accuracy, persistence and real-time performance; if they can only provide demo videos and not reproducible metrics, do not include it in safety arguments yet.
Limits and open questions
As of November 4, 2026, when checking the live KimiSeek catalog, Apollo World is not in the current list, so there is no corresponding KimiSeek access method or price information; the report also did not publish a public API, weights, license or pricing. The deployment mentioned in the report points to Baidu's own Apollo Go scenario; any availability of Baidu APIs refers to Baidu's official services, not KimiSeek availability. If Baidu later announces an API, its quotas and prices should be based on Baidu official information, not external inference. The source is a product event report, lacking technical papers, third-party reproduction and failure cases; the '2+3' criteria are a framework proposed by Baidu, not an industry-standard benchmark. The operating mileage and '100,000 km per day' in the report are statements by Baidu or industry figures, without statistical definitions. Whether Apollo World will be opened to third parties or support robotics scenario migration has no public evidence.
Illustrations from the original report







