L3 License at 120 km/h: More Like Passing a Basic Driving Test

Getting the Beijing L3 test license, raising the speed limit to 120 km/h—the headline for Stelato G9 came precisely. But as an engineer who deals with code daily, what I see isn't the gimmick of being the "first," but the gold content of the testing system behind it, and what this "120 km/h" actually implies.

Here's an analogy. Autonomous driving test licenses are like driver's license exams. Level 3 means "allowed to take hands off the wheel on highways, but must be ready to take over at any time." Beijing's L3 test license review system is roughly equivalent to Subject II—the course exam. You need to perform prescribed actions standardly in closed or semi-closed sections, like lane keeping, automatic lane changes, and handling sudden obstacles. But the condition of 120 km/h upgrades the difficulty of Subject II from parallel parking directly to "continuous S-curves on a highway." Because once speed increases, perception latency, decision tolerance, and control precision amplify exponentially.

Stelato G9's solution is "LiDAR + Vision + Millimeter Wave." This combination isn't new in the industry, but note its LiDAR configuration. If I'm not mistaken, it uses Huawei's 192-line LiDAR, or one with even more lines. Why? At 120 km/h, for an obstacle 150 meters ahead, the system has less than 4.5 seconds to react. The number of LiDAR lines determines point cloud density; more lines mean more reliable identification of distant small targets. In extreme scenarios, like a tire falling onto the highway, vision might misjudge due to lighting or texture, so LiDAR is the last line of defense.

But what truly interests me is the "test license" stage itself. Getting an L3 license in Beijing, with its high-complexity road conditions, means Stelato G9 passed at least thousands of hours of simulation scenario tests and hundreds of kilometers of closed-road tests. Note, the keyword here is "closed roads." Real highways also have ramp merges, construction zones, cars braking suddenly ahead, and even pedestrians crossing (though theoretically not allowed). L3 requires the system to handle these scenarios autonomously, and safely hand over if it can't. And defining "safe handover" is the hardest part to quantify in testing.

From an engineering perspective, I infer Stelato G9's testing focus should be on three levels: First, perception system identification of "static objects." On highways, the most dangerous thing isn't moving cars, but stationary broken-down cars or obstacles, because vision and millimeter-wave radar detection of static targets is easily affected by the Doppler effect. Second, smoothness of lane change strategies. L3 allows drivers to take hands off, but the acceleration/deceleration curves during system lane changes must be smooth and match human expectations, otherwise passengers get motion sickness, and drivers get scared into taking over again. Third, redundant design. In case of single-point failure (e.g., LiDAR blocked by mud/water), can the system switch perception links in milliseconds and degrade to a safe state?

These technical details won't be in the news, but test reports are full of blood, sweat, and tears. Beijing, as one of the strictest cities for autonomous driving testing in China, has a whole cycle for reviewing L3 licenses: application-simulation-closed test-open road test-expert review. Stelato G9 getting it shows it performed adequately in this process. But "adequate" doesn't equal "good to use." I've seen too many cases where performance was perfect in tests but fell apart on real roads. The biggest pitfall of L3 is "ambiguous system boundaries." For example, if the system encounters a sudden downpour at 120 km/h, should it immediately slow down and request takeover, or keep trying? If it requests takeover, the driver might be sleeping or looking at their phone, and a few seconds of takeover time is simply unrealistic.

So, this license is more like a "milestone achievement." It proves Stelato G9 can run through core L3 functions under preset conditions. But there's still a hurdle before true "commercializable L3": how to define and verify "takeover time upon system failure." Current industry consensus is that L3 needs to give drivers at least 10 seconds of buffer. But 10 seconds at 120 km/h means the car travels 333 meters, implying the system must judge its impending failure much earlier and issue warnings in advance. This requires more complex "metacognition" ability—the system knowing what it doesn't know.

Mentioning this image, it likely shows Stelato G9's sensor layout. From the image, we can see forward LiDAR, side blind-spot filling LiDAR, and camera arrays. This "redundant but not over-stuffed" design fits the pragmatic style of engineering. But sensor position and angle directly determine the size of coverage blind spots. For example, if side LiDAR is installed on the fender, detection capability for low objects is affected by body height. Only those who have actually run tests know these details.

Finally, leaving an open question: When an L3 system's takeover request is ignored by the driver, leading to an accident, who is responsible? Is it an algorithm flaw, or human laziness? The answer to this question might be harder to find than getting ten test licenses.

Original link: https://www.ithome.com/0/976/646.htm