For all the hype around China's advanced driver-assistance systems (ADAS), the hardest problem isn't teaching a car to drive—it's teaching it when to stop. Drivers across the country have reported a phenomenon known as "phantom braking," where a vehicle cruising smoothly on an open road suddenly slams on the brakes for no visible reason. The result: seatbelts tighten, phones fly, and trust evaporates.
According to a report by the Chinese outlet Huxiu, this issue is the most common complaint about intelligent driving on forums like X, Reddit, and Xiaohongshu. One Chinese smart-car owner told Huxiu that during a highway lane change, a car passed on the right, and only seconds later did his vehicle finally brake—a hesitation that mirrors complaints about Tesla's FSD, which has been criticized for indecisive lane changes and inappropriate stopping distances at stop signs.
The braking dilemma: safety vs. comfort
Phantom braking stems from a fundamental trade-off. If sensors are tuned for maximum safety, they may interpret a floating plastic bag or a reflection on a manhole cover as a solid obstacle, triggering a harsh stop. If sensitivity is reduced for comfort, the system might miss a real hazard, like a stationary white truck or a pedestrian in dark clothing at night. Human drivers use common sense—a skill machines lack.
Traditional ADAS relied on a three-step pipeline: perception, planning, and control. Engineers manually coded rules for countless corner cases, but this approach couldn't cover every scenario, leading to unpredictable braking. In 2023, Tesla CEO Elon Musk shifted the industry toward "end-to-end" neural networks, which replace hand-coded rules with a single model trained on raw video. Chinese firms like Huawei, Xpeng, Li Auto, and Baidu have since reorganized their ADAS teams around this approach.
New models, new challenges
End-to-end systems excel at generalizing from data, but they introduce two new problems: they are uninterpretable and lack logical reasoning. To address this, researchers are exploring Vision-Language-Action (VLA) models, which add a language-based reasoning layer to help the car understand traffic signs and police gestures. Additionally, "world models"—AI systems that simulate physical environments—are being used as virtual referees to test decisions before they are executed on the road.
Despite these advances, phantom braking persists. A Reddit user quoted by Huxiu described FSD as "a god 99% of the time," but in that 1%, it nearly drove his family into a concrete barrier. He canceled his subscription that night. The trust deficit remains the biggest hurdle for autonomous driving, and until the industry solves the braking problem, the technology's promise will remain unfulfilled.
