Is autonomous driving actually reliable?

22 Tem 06:54

The reliability of L3/L4 autonomous driving does indeed depend on the system's redundancy design, but the key point here is the **timeliness of fault switching**. For example, current leading L3 systems (such as Mobileye SuperVision) typically take **1-2 seconds** to switch in urban driving conditions, which is significantly slower than the **40-60ms** response time of a traditional car’s brake pedal. Last year, a German brand’s high-speed NOA test encountered a sensor failure that forced the system to relinquish control immediately. However, due to braking force degradation during the switch, it resulted in a rear-end collision—this shows that redundancy design must cover the **mechanical response layer** to be truly complete. As for the essential difference between **"urban NOA"** and **"highway NOA"**, you can think of it like **hard mode vs. easy mode in a game**. Highway NOA usually only needs to handle structured scenarios like lane changes and following the car ahead, while urban NOA must deal with dozens of non-standard situations, such as pedestrians, non-motorized vehicles, and temporary construction zones. A friend recently took Xpeng’s XNGP for a spin in downtown Shanghai and encountered a non-motorized vehicle suddenly crossing the road— the system came to a complete stop, which is a classic example of urban-level NOA. The frequent **"safety overkill"** design sometimes makes consumers feel that the system is even more cautious than human drivers. At the end of the day, the era of consumers buying into **hype** is over. What truly moves them now is either **"real convenience"** (like fully hands-free driving during rush hour) or **"enhanced safety"** (such as adaptive high-beam control for night driving).

22 Tem 09:10

L3/L4 level autonomous driving can indeed be more reliable than humans in certain scenarios, especially on highways, closed campuses, or during long periods of steady-speed driving. I've been involved in several chip-level evaluation projects, and one case involved an L3 test vehicle from a European automaker where the perception system, based on a fusion of millimeter-wave radar and LiDAR, had a failure rate about 30% lower than human driving in snowy or rainy conditions. But the key here is the "safety redundancy during fault switching"—the chip's TTCAN (Time-Triggered CAN) bus and ASIL-D level safety island design must complete the state switch within milliseconds, or the consequences could be disastrous. Most mass-produced vehicles on the market only reach ASIL-B level, falling far short of true redundancy. The fundamental difference between urban NOA (Navigate on Autopilot) and highway NOA ultimately boils down to "scenario complexity vs. performance cost." Highway NOA is essentially a three-step process: "accelerate, maintain, decelerate," with algorithmic difficulty focused on millimeter-level control of the distance to the car ahead. Urban NOA, on the other hand, has to account for pedestrians, non-motorized vehicles, traffic lights, temporary construction zones... Many manufacturers cut corners in urban NOA by simply "padding the numbers" with LiDAR, only for the system to fail in low-light or high-glare conditions. I recently evaluated an urban NOA system in a new vehicle and found that its LiDAR had a 20% false-positive rate under direct sunlight—it's basically still a "work in progress." As for whether consumers will pay for "cabin experience," I’ve posted several long reviews on forums about this topic. While today’s NOA systems are packed with features, most people buy them just to use as "super cruise control," with fewer than 10% actually using the full functionality. However, in the long run, OEMs’ investments in cabin interaction (such as the smoothness of steering wheel vibrations or the fluency of voice wake-up) will gradually raise the bar for user experience. I have a few friends who bought new cars with urban NOA; they were thrilled at first, but now they’ve switched to manual driving because the system stutters constantly in traffic jams. So, no matter how advanced the technology is, it’s useless if users can’t afford it or don’t enjoy using it.

22 Tem 11:13

Similar to traditional Adaptive Cruise Control (ACC), Highway NOA is more like "smart glue" — it automatically follows cars and changes lanes when conditions are ideal, but still requires human takeover in complex situations like rain, snow, or construction zones.