Baidu launches Apollo World to simulate robotaxi roads

On Tuesday, November 3, Baidu launched Apollo World at its AI Day in Shenzhen. It is a world model for physical Super Intelligence: a system that builds a picture of how a road environment works, then predicts what happens next. The practical point is simple. Apollo Go robotaxis can practise inside the simulation, including situations the fleet rarely meets on real roads.

Baidu's headline number is easy to picture. Apollo World can generate up to 100,000 kilometres of simulated driving a day, which Baidu says equals about three years of real-world testing by one safety-driver vehicle. The system produces real-time multimodal data streams, such as camera and lidar observations, so a self-driving system can sense and navigate the simulated environment directly, rather than watch a recording of it.

Agents inside the simulation, not just scenery

What sets Apollo World apart from a digital map is how it treats the people and vehicles on the road. Vehicles and pedestrians inside the simulation act as agents: each one makes its own decisions and reacts to the others. So a robotaxi is not memorising one script. It is dealing with traffic that changes because other participants made choices. The foundation is real driving: Baidu says the model draws on more than 350 million kilometres of Apollo Go testing, then stretches that history into situations the fleet has never encountered.

Baidu also proposed a new way to grade world models, called the 2+3 standard. The reports we read do not spell out what the 2 and the 3 stand for, so we will not pretend to explain them. What the announcement does make clear is that Baidu wants a say in how this category is judged, not only in how it is built. The company also says the stack can extend to robotics and other physical Super Intelligence uses, though for now its first user remains Apollo Go.

Why simulation is a big bet

The context is cost. Building self-driving technology stays expensive, even though many Chinese robotaxi firms expand abroad asset-light, supplying technology and partnering with automakers and local fleet operators instead of owning every fleet outright. Reuters quotes mobility expert Frank McCleary making the point that launch speed comes down to how fast the technology adapts to a new environment. Simulation is Baidu's answer to that problem: more learning without adding vehicles and safety drivers at the same rate.

Apollo Go is already China's largest robotaxi operator and has been tested globally, competing with WeRide, Pony.ai, and Momenta Global. Morgan Stanley has estimated China's robotaxi fleet could reach 1 million vehicles by 2035, and XPeng, Volkswagen's Chinese partner, plans to start robotaxi operations in 2027. Those are a projection and a plan, not promises. What is real today is narrower: Baidu now trains its robotaxis in a world it built itself, and that world produces 100,000 kilometres of practice a day. Meituan's coding model for long jobs is in our LongCat article. The new test where even the best model scored only 12% is in our GDPx article.

Sources: Reuters (Nov 3, 2026), Crypto Briefing (Nov 3, 2026). Figures and status are snapshots at writing time, evening November 3, 2026 WIB, and can change.

Written by Mastechon

Writing about crypto and Super Intelligence: market news, new projects, Super Intelligence models, and early opportunities like airdrops and testnets. Every piece starts from the official announcement, with the costs worked out up front.