The secret behind Radish Run's "stall": the autonomous driving network got "stuck"

On March 31, 2026 at 8:57 p.m., on Wuhan’s Third Ring Road, a Robotaxi—an autonomous vehicle from Baidu’s subsidiary Luobo Kuaipao—suddenly slowed down. It then came to a complete stop in the center of the fast lane. Ten minutes later, not just this one vehicle—on Taizi Lake Bridge, Baishazhou Bridge, the Second Ring Road, and Yangsipu Bridge—nearly 100 white autonomous cars almost simultaneously “lost their minds,” stopping on the main thoroughfares.

A passenger said they were trapped on the elevated road for nearly two hours. The vehicle’s SOS button wouldn’t connect, the customer service line was busy, and in the end, traffic police had to walk up to the elevated road to rescue the vehicles one by one.

The next day, Wuhan traffic police issued a report, initially concluding that the incident was caused by a system malfunction.

In Wuhan, locals have given Luobo Kuaipao autonomous vehicles a nickname—“Shaoluobo.” In Wuhan dialect, “Sha” means “stupid.” This was not Luobo Kuaipao’s first time “breaking down.” In July 2024, there were also vehicles that suddenly stopped on the road during the evening rush hour in Wuhan. In the end, after calling traffic police’s customer service, the vehicle occupants got into the driver’s cabin and drove the car to the roadside. With the same “system malfunction,” and the same “traffic police stepping in as a safety net,” nearly 2 years later, the problem appeared again.

Luobo Kuaipao is Baidu’s core business in the intelligent driving and mobility services sector, and is an important direction for Baidu to commercialize its AI technology. According to disclosures in Baidu’s 2025 annual report, the operating entity of Luobo Kuaipao is Luobo Yunli (Beijing) Technology Co., Ltd. This company is 100% wholly owned by Apollo Intelligent Technology (Beijing) Co., Ltd. (Baidu subsidiary; abbreviated as “Apollo Apollo”), which is under Baidu.

As of now, both Baidu and Luobo Yunli have not publicly explained the cause of this malfunction. Caixin Finance contacted Baidu; the company said that no related information is available to disclose yet.

In a research report released by Guidehouse Insights in the fourth quarter of 2025, Baidu Apollo is listed as one of the two global leaders in the autonomous driving field, with the other being the U.S. company Waymo.

According to the description, Luobo Kuaipao’s autonomous driving technology is supported by Baidu’s four-layer full-stack AI architecture (cloud infrastructure, deep learning frameworks, large models, and applications).

As of February 2026, its global rollout has expanded to 26 cities. Since February 2025, in all operating cities on the Chinese mainland—including Beijing, Shanghai, Shenzhen, Wuhan, Chengdu, Chongqing, Haikou, Sanya, and others—it has achieved 100% fully driverless operations (i.e., no safety staff inside the vehicle), corresponding to L4 level, and has obtained paid-service permits in multiple cities. The cumulative number of autonomous driving mobility service orders provided to the public has exceeded 20 million.

In international markets, it has also entered cities such as Dubai, Abu Dhabi, London, Saint Gallen, and Seoul. On March 30, Luobo Kuaipao officially launched fully driverless commercial operations in Dubai. But the very next day, Wuhan saw a scene of a collective “breakdown.”

What is the cause of this malfunction, and can the problem be resolved? Caixin Finance interviewed Zhu Sichan, a professor at Tongji University’s School of Automotive and Energy, and the chief expert on intelligent connected vehicles at the China Automotive Engineering Research Institute. He has long focused on automotive safety and intelligentization research, and is one of the main builders of my country’s first set of automobile collision safety standard system. In recent years, he has focused on intelligent driving testing and evaluation and standard-setting. As the head of the i-VISTA China Intelligent Vehicle Index evaluation system, he has deep insights into intelligent driving technology routes, safety bottom lines, and standard systems.

Zhu Sichan has repeatedly issued calm commentary on industry booms. He has stated directly, “If someone announces end-to-end mass production on the road, you can’t buy this car,” and he advocates a fusion architecture of “raise the ceiling for AI, set the rules to guarantee the bottom line.” He has also been critical of the “10-second takeover” rule for L3-level autonomous driving, believing that safety must be addressed step by step, all at once.

The following is the interview:

Caixin Finance: Wuhan traffic police initially judged that the cause of this incident was a ‘system malfunction.’ From your professional perspective, what technical problems does this situation expose in autonomous driving?

Zhu Sichan: Many people think autonomous driving is “the car drives itself,” like a human driver—one driver, one car, and everyone drives freely. That’s called “single-vehicle intelligence.” The Robotaxis operating globally are not on that track.

They follow the “connected intelligence” route. At present, in Robotaxi vehicles, the driver position has no safety personnel, but in the cloud-based supervision platform in the back end, there are safety supervision personnel. I went to visit the cloud supervision platform at the intelligent connected vehicle demonstration zone. The supervisors sit in a device similar to a driving simulator. On a large front screen, the running autonomous vehicles are displayed. If problems arise, supervisors can intervene remotely to solve issues that the autonomous-driving vehicle-side software cannot resolve.

Right now, the Robotaxi falls under the category of connected intelligence. Simply put, it’s five things working together: vehicle, road, cloud, network, and map.

Vehicle: The most important is the intelligent vehicle. It has multi-environment perception sensors made up of cameras + millimeter-wave radar + LiDAR, supported by AI compute power from high-compute chips. It uses an automated driving system built with a visual language large model (VLA) + an end-to-end deep learning model;

Road: Smart roads, collecting road traffic information, with real-time traffic conditions on the map, used for path decision-making of autonomous driving vehicles;

Cloud—what I mentioned earlier—has back-end platform personnel supervising. In China, in most Robotaxi demonstration operations, one supervising staff member can supervise 5–10 driverless vehicles. In the United States, Waymo also has back-end supervisory personnel, but I haven’t seen the parameter for how many vehicles each person supervises;

Network: The communication network between the vehicle and the cloud;

Map: navigation maps, high-precision maps, and fused positioning resources including satellite positioning, ground base-station positioning, vehicle inertial navigation, and others for integrated localization.

These five are all indispensable. If any one link has an issue, the system will trigger a safety mechanism—stopping. On the evening of March 31, the problem may well have been “the network.”

Caixin Finance: So the problem isn’t in the ‘vehicle,’ but in the ‘communications’—right?

Zhu Sichan: Yes. There is a publicly known “industry secret” here: when connected intelligence technology was proposed, it also proposed building a 5G communications network. By constructing 5G base stations, it would achieve a real-time communications network with high reliability, super bandwidth, and low latency. However, in the early construction of intelligent connected vehicle demonstration zones, 5G communications technology was not successful. The communications network used by Robotaxi operations is not the legendary 5G communications; it is still the consumer-grade network that we use on our phones. Stuttering and dropped connections cannot be avoided.

The fault on the evening of March 31 likely resulted from the communications network failing. The link between the vehicle and the cloud was cut. The system triggered a safety mechanism, and vehicles within the network-fault area collectively stopped. This is not that the “vehicle is broken,” but that “the network is down.”

Caixin Finance: Then the vehicles directly stopped in the middle of the roadway lane, rather than pulling over—was that a design flaw?

Zhu Sichan: First, I need to affirm one thing: Luobo Kuaipao has safety fallback measures. When communications are cut, the vehicle stops; it does not let the vehicle continue “running totally unprotected.” In terms of safety logic, that is correct. But we are not satisfied with the control logic of the safety measures.

The vehicle-side hardware and software did not fail. It has a high-precision map, knows it is on an elevated road, and also knows there are lanes next to it, exits, and roadside areas. As an L4-level autonomous driving system, it has the ability to change lanes autonomously.

So what is the reasonable approach? After triggering the safety mechanism, it should drive away from the urban expressway and pull over to the roadside to park, and inform passengers to get out of the vehicle. After reaching a safe area, it should then contact the supervision platform again.

But now, it stopped directly in the fast lane. That’s why locals in Wuhan gave it the nickname “Shaoluobo”—in Wuhan dialect, “Sha” means “stupid,” and it’s well deserved.

Of course, we should have some tolerance for the development of artificial intelligence technology. When problems are found in Robotaxi demonstration operations, technology companies need to iterate quickly, so that the technology can mature rapidly and benefit society.

Caixin Finance: Many passengers report that the SOS wouldn’t connect, the customer service line was busy, and ground staff didn’t arrive for a long time—what’s the reason?

Zhu Sichan: If the communications network is normal, the cloud-end supervision personnel can handle things quickly in the background through remote control, so the impact on road traffic is not big. But when communications are cut, the background cannot be handled remotely. If the back-end cloud supervision platform cannot intervene, it can only rely on ground personnel—i.e., “ground staff”—to handle it on site. Some of the “breakdown” vehicles that night were on elevated roads, which would itself trigger congestion. Ground staff being stuck in traffic makes it hard to arrive in time, and typically it takes a long time.

Based on this information, Luobo Kuaipao’s operation in Wuhan complies with the management rules of the intelligent connected vehicle demonstration zone. This is not a problem only affecting Luobo Kuaipao. Four months ago, the entire city of San Francisco had a power outage, and Waymo’s driverless cars also stopped collectively at intersections and turned on their hazard lights.

Caixin Finance: Then for the legal responsibility of this ‘breakdown’ incident, who should bear it?

Zhu Sichan: When it comes to responsibility and legal gaps, these operational Luobo Kuaipao driverless taxis are vehicles approved and recognized by the Wuhan intelligent connected vehicle demonstration zone for road demonstration operations. At present, the responsibility and legal provisions for traditional vehicles and road traffic cannot be fully applied; the determination should be made based on the management rules of the intelligent connected vehicle demonstration zone.

Caixin Finance: Baidu’s annual report shows that Luobo Kuaipao has been operating in 26 cities nationwide, with cumulative orders exceeding 20 million. What do you think about the relationship between this expansion speed and public safety?

Zhu Sichan: Demonstration operations through intelligent connected vehicle demonstration zones are meant to expose problems, fully surface the issues, and make timely improvements—only then can the product mature.

Autonomous driving cars are the most valuable application scenario for artificial intelligence and play a major role in the digital economy. You cannot kill technological exploration and development because of one or two incidents. At the same time, autonomous driving car safety requirements for testing technologies, evaluation standards, and accepted certification methods are also being actively advanced. Industry participants should each do their part; technology companies work to iterate products; research institutes innovate evaluation methods for intelligent agents; and government authorities explore innovative approaches to safety regulation of intelligent agents.

Safety problems need to be addressed, and progress in technology should be met with tolerance. We can’t let safety hazards of new technologies run wild, but we also shouldn’t call for a crackdown at the first sign of trouble and suffocate development of new technologies.

Caixin Finance: What about safety redundancy?

Zhu Sichan: Of course, we also can’t ignore this: the pace of building safety redundancies for Luobo Kuaipao seems to lag behind the pace of commercial expansion. But is this just Baidu’s problem? Not entirely. However, right now there are still relatively few industry participants, so you can’t say this is “a common problem throughout the industry.”

Caixin Finance: What do you think Luobo Kuaipao should do next?

Zhu Sichan: Based on this incident, I have three recommendations.

First, Luobo Kuaipao’s company should improve its public relations strategy. When problems occur, it should release official updates in a timely manner to explain the technical nature of the incident. Only through transparent information releases can social sentiment be prevented from growing indefinitely.

Second, optimize the “minimum risk strategy.” Based on current information, after Luobo Kuaipao triggers a safety mechanism, its “minimum risk handling strategy” should still be optimized. After diagnosing a communications network failure, it is reasonable to stop normal operation of driverless vehicles in the affected network-fault area. However, having an intelligent vehicle drive by itself away from an urban expressway, find a safe roadside spot to park, and guide passengers to get off the vehicle—this more reasonable minimum risk handling strategy should be incorporated into autonomous driving algorithms.

Finally, continue exploring more reliable communications solutions. In “vehicle-road-cloud-network-map” (the ‘network’ part), the weakest link is currently the “network.” Even if early construction of vehicle-class 5G communications foundations for “connected intelligence” faced obstacles, communications technology should still be explored and improved within connected-intelligence solutions for autonomous driving, to pursue more reliable communications solutions for vehicle-to-cloud interconnection in “vehicle-road-cloud-network-map.”

Caixin Finance: My final question—if your family asks, ‘Can I take Luobo Kuaipao now?’ what would you say?

Zhu Sichan: My attitude is: you can try it for daily commuting, but be psychologically prepared—it’s still not smart enough.

(Editor: Guo Jiandong ）

     【Disclaimer】This article only represents the author’s personal viewpoints and is not related to Hexun.com. Hexun.com maintains a neutral stance toward the statements and judgments in the text, and does not provide any explicit or implied guarantees regarding the accuracy, reliability, or completeness of the content included. Readers should use this information for reference only and bear all responsibility themselves. Email: news_center@staff.hexun.com