Baidu Unveils Apollo World Model for the First Time; Apollo Go's End-to-End System Now in Commercial Operation
nashnova research
Baidu disclosed Apollo World, a world model for autonomous-driving simulation, and revealed that roughly 100 Luobo Kuaipao robotaxis running its new end-to-end architecture have entered live passenger service — with fully driverless mileage approaching one million kilometers — marking the shift from closed testing to real commercial operation.
What does the end-to-end architecture actually change?
Traditional self-driving stacks split the job into modules — perception, prediction, planning — each governed by hand-coded rules. End-to-end architecture (a single model that takes raw sensor data and outputs steering, throttle, and brake commands directly) lets the model learn to drive from massive driving data.
This means → the development logic flips from "engineers write rules" to "data trains the model," mirroring how large language models predict the next word — except the prediction target becomes steering angle, throttle, and brake.
In plain terms = engineers used to tell the car "stop at a red light." Now the car watches millions of kilometers of video and figures that out itself. The upside: a higher capability ceiling. The risk: the car may learn behaviors engineers never anticipated, making safety verification far harder.
About 100 cars on the road — is the data strong enough?
Baidu IDG chief architect Wang Liang disclosed that roughly 100 Luobo Kuaipao robotaxis now run the end-to-end architecture in live passenger service, with fully driverless safe mileage approaching one million kilometers.
This means → the architecture is no longer a closed-track experiment — it is picking up and dropping off real passengers, crossing from technical validation into commercial service.
Broader context: as of Q2, Luobo Kuaipao had provided over 23 million rides, logged more than 350 million km of autonomous driving — of which over 240 million km were fully driverless — across 28 cities worldwide.
What problem does Apollo World solve?
End-to-end models can produce unexpected behaviors. Traditional simulation — replaying recorded road footage — cannot answer "what if the oncoming car reacted differently?" Apollo World's core shift: it models every vehicle and pedestrian in the virtual environment as an AI agent capable of independent decisions, not a scripted replay.
In the live demo, at the same intersection the robotaxi braked when the oncoming car swerved around it, but drove straight through when the oncoming car stopped to wait — change the other actors' behavior, and the robotaxi's decision changes accordingly.
Baidu says Apollo World can run roughly 100,000 km of virtual road-testing per day. But Yang Ze, head of the world-model effort, acknowledged that the model still cannot fully simulate every variation in the real world — virtual data must be combined with real road data.
"Safety doesn't mean being conservative" — what tension sits behind that line?
Wang Liang said that as Luobo Kuaipao's fleet grows, robotaxis interact with human drivers more often; an overly cautious strategy actually disrupts traffic flow and degrades the passenger experience — the cars need to blend into local traffic.
In plain terms = a robotaxi that is always half a beat slow becomes a "mobile roadblock" — safe for itself, a bottleneck for everyone else.
A second challenge: not all human driving behavior is worth learning from. Irregular habits can "pollute" training data. Baidu must filter aggressively, and adapt to each city — Chongqing's multi-level interchanges, Hong Kong's and London's left-hand traffic — completely different traffic rhythms.
Could general-purpose large models just take over?
World models are an industry-wide pursuit: Waymo released its Waymo World Model in February; Wayve shipped GAIA-4 in August. In China, Pony.ai launched PonyWorld 2.0 in April; WeRide released WeRide GENESIS in January.
The bigger wildcard is general-purpose large models. In September the independent DrivingBench project used GPT-6 Astra and three other foundation models to control a real Toyota Corolla at low speed. GPT-6 Astra completed the full cone course on its second attempt — 134.7 meters, averaging about 1.5 km/h — but the other three models failed to pass the halfway mark, and a safety driver was ready to intervene throughout.
This means → foundation models can now "drive a short stretch" at very low speed under human supervision. But Yang Ze noted that autonomous-driving systems must make decisions every ~33 milliseconds — foundation models remain far from handling real L4 tasks. Whether Apollo World can genuinely lift end-to-end safety-verification efficiency will be the key test of this technology stack's real-world value.
市场有风险，内容仅供研究参考，不构成投资建议。