Apollo Go Explained: From Origins to Now

Why it is suddenly everywhere

1. 01

   Baidu’s August 2026 results put Apollo Go in 28 cities, with more than 350 million autonomous kilometres accumulated—over 240 million of them fully driverless.

2. 02

   The same first-party update says London open-road tests involve Uber and Lyft, Dubai has fully driverless commercial operations, and Hong Kong has begun fully driverless testing in a right-hand-drive, left-traffic market.

3. 03

   Wuhan supplied the scale model: the city government reported 3,379 kilometres of open test roads, roughly 3,000 square kilometres of coverage and more than 2.5 million autonomous-mobility orders by early 2025.

4. 04

   Scale also exposed a new risk. Wuhan police said a system fault stopped multiple Apollo Go vehicles on roads on March 31, 2026; passengers exited safely and nobody was injured.

Apollo Go—known in Chinese as Luobo Kuaipao, or “Carrot Run”—is Baidu’s autonomous ride-hailing service. A rider requests a car through an app, but the visible absence of a driver is only the last link in a much larger system: mapped operating zones, sensor-heavy vehicles, remote support, depots, regulators and city infrastructure all have to work together.

That distinction explains why the service matters now. Robotaxis spent years as impressive demonstrations contained inside carefully chosen districts. By late 2026, Apollo Go was no longer presenting one Chinese pilot as proof. It was testing whether a system refined in Wuhan could cross borders, traffic conventions and regulatory regimes. London, Dubai, Hong Kong and Switzerland are not simply extra pins on a map; together they ask whether the operating playbook travels. The same question now faces the Western front-runner, with Waymo testing its playbook against a Toronto winter.

The timeline

1. 2013–2017Research becomes a platform

   Baidu begins autonomous-driving research in 2013 and introduces the Apollo open platform in 2017, shifting from an internal car project toward a developer and industry ecosystem.

2. 2020Robotaxis meet the public

   Baidu’s Q3 report says Apollo Go opens to the public in Beijing after Changsha and Cangzhou, with rides ordered through Baidu Maps or the Apollo Go app.

3. 2022The driver leaves the cabin

   Wuhan approves Apollo Go’s street rollout; in Shanghai, Baidu starts fully driverless testing and reports that Wuhan and Chongqing already offer commercial rides without an onboard safety operator.

4. 2024Wuhan becomes the proving ground

   Cross-river service and wider operating zones make the cars part of daily urban life. Low promotional fares, tourist curiosity and driver anxiety push the service into national debate.

5. 2025A city-scale system takes shape

   Wuhan reports 3,379 kilometres of open roads and an autonomous-mobility service area of about 3,000 square kilometres; remote operations and road infrastructure become as important as the vehicle.

6. 2026-03A fleet-level failure

   A system fault stops multiple vehicles in Wuhan. No one is injured, but the incident reveals that centralised software and connectivity can turn one fault into a city-scale event.

7. 2026-08The playbook goes abroad

   Baidu reports a 28-city footprint, open-road tests in London and Switzerland, fully driverless commercial operations in Dubai, and fully driverless testing in Hong Kong.

It began as infrastructure, not a taxi app

The easiest way to misunderstand Apollo Go is to see it as a ride-hailing app with the driver removed. Its deeper origin is Apollo, the autonomous-driving stack Baidu opened to partners in 2017 after starting research in 2013. The stack had to combine perception, prediction, planning, localisation and vehicle control. Public rides were therefore not the beginning of the technology; they were the moment a long-running engineering platform acquired a consumer-facing service.

By the third quarter of 2020, Baidu described Beijing as Apollo Go’s third public city after Changsha and Cangzhou. That sequence matters. The early strategy was not to declare a universal autonomous car, but to learn one bounded operating domain at a time. Each city supplied different road geometry, weather, driving behaviour and regulatory relationships. Every expansion created data, while every restriction made the limits of the system explicit.

The name change also did cultural work. “Apollo” sounded like an engineering programme. “Luobo Kuaipao” sounded local, playful and ordinary enough to hail after work. The product breakthrough was partly this conversion of an abstract AI system into a repeatable urban service.

Why Wuhan changed the story

Wuhan offered three things at once: a large and complicated road network, an established car industry and policy willingness to open roads. Rivers divide the city, so a service that could cross bridges was visibly more useful than one circling an isolated development zone. The city’s January 2025 account said its open test network had reached 3,379 kilometres across 12 districts, covering around 3,000 square kilometres. That scale made Wuhan less like a laboratory and more like an operating city.

It also made the service socially legible. In 2024, local drivers interviewed by ChinaVenture described the cars as cautious, easy to cut in front of and unable to enter some complex transport hubs. Riders, meanwhile, were attracted by privacy, novelty and heavily discounted fares. Apollo Go even became a tourist activity. These apparently contradictory impressions—too clumsy to replace a driver, yet credible enough to worry drivers—were the real sign of a breakout. The technology had moved from “Can it drive?” to “Who benefits if it can?”

Cost was the other turning point. Securities Times reported that the sixth-generation vehicle’s announced price was RMB 204,700, 60% below its predecessor, with Baidu targeting 1,000 units in Wuhan. Those were company plans rather than proof of lasting profitability, and promotional fares obscured the underlying economics. Still, cheaper purpose-built vehicles changed the question from whether robotaxis were technically possible to whether utilisation, maintenance and remote support could make them economical.

The “driverless” system still employs drivers

Removing a person from the front seat does not remove people from the loop. Sixth Tone’s reporting on Wuhan safety operators described two layers of work. In-vehicle testers intervene during development and record difficult scenarios; remote operators monitor cars in service and issue instructions when a vehicle cannot safely resolve an edge case. Roadworks, aggressive merging, unusual cargo and even sprinkler spray become training examples.

This human layer is not an embarrassing exception to autonomy. It is part of how a bounded L4 service works: the vehicle drives itself inside specified conditions, while an operations organisation handles exceptions. Depots clean and charge cars. Mapping teams update roads. Remote staff triage problems. Regulators define where and when vehicles may operate. “Driverless” describes the cabin, not the whole labour system.

That creates a more complicated employment story than either “robots eliminate drivers” or “technology creates new jobs.” A former ride-hailing driver interviewed by Sixth Tone said safety work offered similar income, insurance and more stable hours. But ChinaVenture noted that new roles in testing, maintenance and remote support may not match the number of driving jobs exposed to automation. The distribution of work changes, as do the skills and institutions around it.

The outage that revealed the network

On March 31, 2026, multiple Apollo Go cars stopped in Wuhan roads. According to the police statement relayed by China National Radio, calls began reaching emergency services at 20:57. Traffic and transport authorities responded with company staff; passengers were brought out safely, there were no injuries, and the preliminary cause was a system fault.

No collision occurred, which supports the value of a fail-safe stop. Yet a stationary vehicle can still create danger and congestion, especially when many stop together. The event exposed the reverse side of fleet intelligence. Centralised maps, software and communications let lessons spread quickly across every car, but they can also create correlated failure: a weakness shared by the fleet rather than isolated to one driver. That tension is becoming a wider infrastructure question as the bridge from 5G-A to 6G promises stronger uplinks, sensing and network intelligence.

That is why robotaxi safety cannot be reduced to kilometres between crashes. Cities also need to ask how quickly passengers can exit, how vehicles clear live lanes, whether remote systems degrade gracefully, and who coordinates recovery when many cars fail simultaneously. The Wuhan incident did not settle whether autonomous driving is safe. It made the safety question more precise.

Why the global expansion is the real test

Baidu’s August 2026 update described a footprint of 28 cities and more than 350 million autonomous kilometres. More importantly, each new market changes a foundational assumption. Hong Kong reverses the traffic side used in mainland China. London brings dense streets and partnerships with established ride platforms. Dubai moves from testing to fully driverless commercial service. Switzerland adds another regulatory and road environment. Kazakhstan extends the partnership map into Central Asia.

A system can accumulate enormous mileage by repeating familiar routes. Internationalisation asks whether its methods generalise. Local traffic law must be encoded; maps and pickup behaviour must be rebuilt; emergency procedures need local institutional partners; and riders must trust an unfamiliar service. The next chapter is therefore not a simple race for the largest fleet. It is a test of whether the invisible organisation around the car is portable.

Chinese consumer AI has run into a version of the same wall from the opposite direction. ByteDance’s agent-operated phone was restricted by the very apps it needed to work with — in both stories, the technology was ready before the institutions around it agreed to let it run.

Key facts

Public service

2020

2026 footprint

28 cities

Autonomous distance

350M+ km

Fully driverless

240M+ km

Wuhan open roads

3,379 km

RT6 announced price

RMB 204,700

What people are saying

Common questions

Is Apollo Go truly driverless?

Many commercial rides have no safety operator inside the car, but the service still relies on remote operators, mapped operating domains, maintenance crews and city agencies. “Driverless” accurately describes the cabin; it does not mean human-free operations.

Why did Wuhan become the flagship city?

Wuhan combined extensive open roads, a strong automotive base, supportive rules and unusually varied geography. Crossing rivers and expanding across districts let the system graduate from a fenced demonstration to a service that residents could use in ordinary life.

Did the March 2026 outage prove robotaxis are unsafe?

It proved that fleet-level failures need their own safety plans. The vehicles stopped and nobody was injured, but simultaneous stops obstructed roads and required a coordinated response. Collision safety, graceful degradation and recovery capacity must all be measured.

Will robotaxis immediately replace ride-hailing drivers?

No. Operating zones, fleet size and edge cases still limit near-term substitution. The longer-term pressure is real, however, because lower vehicle costs and around-the-clock operation change industry economics. New safety and maintenance roles exist, but not necessarily at the same scale or for the same workers.