In the blink of an eye, it is the season when Beijing willows are flying in the sky. Whenever this happens, I think of that classic line.
"Spring is coming, everything is reviving, the animals of the savannah..."
I don't know exactly what happened to the animals on the savannah, but I know that the spring of the car circle is indeed coming.
This week can be said to be the most lively week in the car circle since 2022, not only the release of new cars, but also the concentrated appearance of various intelligent driving technologies.
As the first echelon of domestic intelligent driving AI chips, Horizon also chose to bring their mass-produced intelligent driving technology experience day activities at this time.
Nine months ago, we said that Horizon is still an unfamiliar business to many readers; nine months later, I speculated from the number of followers on Horizon's official Weibo that this strangeness may continue for the average consumer.
But in fact, Horizon has launched three production-level intelligent driving solutions, including horizon matrix mono front-view assisted driving program, Horizon Matrix Pilot pilot assistance program, and Horizon Halo in-vehicle intelligent interaction program. And my experience will also unfold in this order.
Mono: One camera to get the L2
Mono, which relies on a single camera to achieve L2-level ADAS-aware performance, is horizon's earliest and most mature forward-looking assisted driving solution, and has been installed in several production models. The Mono experience was also completed on a mass-produced version of the Chang'an UNI-V.
As soon as I got on the bus, before I could fasten my seat belt, the horizon staff first set up a big flag for me, and suddenly raised my expectations several steps.
"It's so easy to run the L2 at high speed, so let's experience Mono's L2 on the city road."
After saying that, this Changan UNI-V really merged into the traffic near the horizon headquarters.
A little disappointed is that the L2 I originally thought was that the full-speed domain ACC adaptive cruise + LKA lane remained open at the same time, which may be limited to the state of the vehicle itself, and the Mono experience only turned on the full-speed domain ACC function throughout the whole process.
But that being said, Mono's full-speed domain ACC is really top.
We're coming in near 5 p.m., the evening rush hour has begun, and the situation on the roads has become more complicated. Private cars, buses, motorcycles, electric bicycles, pedestrians... We have almost all of the traffic participants you can see on the road surfaces of the evening peak city. But it was in this condition that we drove mono, stopping and going through more than a dozen traffic light intersections, while the ACC indicator on the dashboard remained on.
One of the things that reassuring me about Mono is that it visually marks the distance to the vehicle in front of it.
When the reduced distance in front of the car shown on the dashboard came along with the drag of the vehicle's brakes, I knew that Mono was still in steady control and I could just continue to sit.
Of course, Mono also has room for improvement.
For example, perhaps limited by the computing power of the chip, I observed that when Mono recognized a motorcycle, it did not display the motorcycle model on the dashboard at the first time, but first recognized it as a car model, and then switched to a motorcycle model after about 1 second.
However, this is only a look and feel displeasure, and has no impact on its function of controlling the speed of the car.
Pilot: Pure vision enables high-speed NOA
If mono's experience is just the "appetizer" of this event, then pilot's experience is undoubtedly the "core hard dish" of this event.
Before the event began, I posted a photo of "Wild Ideal ONE" on Weibo, which is a horizon research and development test vehicle, and the additional lidar sensor on the car has attracted the attention of many netizens.
But in fact, Pilot is a pure visual solution, which relies on 6 cameras + high-precision maps to achieve high-speed pilot NOA function. The use of lidar here is only "truth calibration" to verify that the results of Pilot's visual perception are accurate, and when you arrive at the production car, you will not see lidar.
Since it is a high-speed NOA, it is natural to go on the high-speed to show strength.
On the high speed, the staff of horizon flicked the gear shift lever, the voice prompt NOA function in the car was turned on, and the HMI man-machine display interface installed by horizon was also changed from a 2D map navigation interface to a 3D real-time lane view, showing the lanes and vehicle information around our vehicles at 360 °.
In general, Pilot can realize automatic follow-up, automatic lane change, actively surpass slow cars, actively offset large trucks, drive in large curvature curves, and automatically enter and exit ramps.
Specific to the subdivision scenario, in everyone's imagination, it may be felt that the driving of the ramp with large curvature is a major problem faced by high-speed NOA.
But it's not. For a mature set of pure visual NOA solutions, as long as the lane lines on the road surface are clear, the driving difficulty of large curvature and sharp turns is not high. Even if you do encounter a corner with a small radius, the system will automatically control the vehicle to slow down and pass calmly.
In contrast, ramp Y-turns with unclear lane lines pose a greater challenge to noa systems.
The difficulty of this scenario lies in the sudden change of lane line.
Many ramps in front of Y-shaped forks are single lanes, and as you approach the forks, the distance between the lane lines on both sides gradually widens until you reach the Y-junction. Approaching a Y-type intersection, the single-lane ramp does not draw an additional lane line in the middle of the road in advance to distinguish between two lanes in different directions.
In our eyes, a small road is divided into two directions, which is clear at a glance and easy to distinguish. But for computers, it relies more on lane line information to determine the direction of progress. In this scenario, the lane line is only widened, not changed, and it is difficult for the computer to judge the road ahead. In many cases, the system assumes that the lane has not changed, but when one lane of the ramp splits into two lanes in different directions, the consequence of this judgment is that the vehicle will crash straight into the guardrail at the fork.
The Horizon Pilot's performance in this scenario is amazing.
After identifying the Y-turnout, Pilot will control the vehicle early in the forward direction to deflect a very slight angle, and the car will remain in a position near the middle of the lane, but I can feel that the car will turn in the intended direction.
This subtle operation allows me to clearly predict the next action of the vehicle, and this pre-judgment brings me a full sense of security and trust.
As for the large curvature of the ramp that everyone had found difficult before, let's say that even in the low visual scene of the direct light of the spring sunset, the Pilot only shook the steering wheel slightly, and then immediately controlled the speed and continued to drive smoothly.
If I had to pick fault with Horizon's high-speed NOA, all I could think of was two.
First, the system is still too conservative in judging the distance between the front and back of the target lane when changing lanes. According to my visual estimates, it can complete the lane change relatively smoothly with the distance between the front and rear vehicles greater than 20 meters.
Second, the system is currently unable to pass through toll booths autonomously. Horizon gives the explanation that this is for safety reasons. Due to the sudden increase in toll booth lanes and the more complex road conditions, the driver is required to take over the vehicle himself before entering the toll booth.
I believe this is also a problem that will be solved in the evolution of Horizon Pilot in the future.
Halo: Multimode interaction
Whether it is Mono or Pilot, behind it is horizon's absolute ability to perceive AI. With this premise, it is natural for Horizon to create a smart vehicle interaction solution Halo.
Halo's core selling point is full-time wake-free.
At present, mainstream voice interaction often requires a "hello, XX" wake word to activate the voice system. For the system, this specific word activation can avoid the problem of misidentification in daily conversation, but it also brings the problem of unnatural interaction. Especially when there are other passengers in the car, suddenly shouting "Hello, XX", I personally feel a little ashamed.
The highlight of Halo's full-time wake-free is that it automatically tells you whether you're speaking to it or to other passengers in the car based on semantics.
"For example, we feel a little hot now and want to open the car window, these four words are a clear instruction, but I am not saying to Halo, Halo also recognized this scene, so our car window was not opened." 」 But when I want Halo to execute my instructions, like this, open the window, you see, the window is open."
The Horizon staff used this example to show me the halo's wake-free features at full time.
It was truly an amazing experience. As human beings, we can tell whether the object of the sentence is ourselves based on the tone of the speech, the size of the voice, and even the orientation. But for computers, the recognition results of different tones of "open window" are actually not much different, and Halo has done exactly this with AI perception.
Because of the labels of "autopilot" and "intelligent driving", Horizon is often compared to Tesla. But in fact, they are not competitors, and if you do the math, the track between them is even separated by several layers.
Therefore, despite making three sets of highly usable intelligent driving solutions, throughout the event, Horizon has been emphasizing that they are a chip manufacturer positioning Tier2. The purpose of making these three sets of plans is also to express to automotive OEMs that their chip platforms can achieve such functions.
Compared with the implementation of these terminal functions, Horizon said that they have made more efforts in the back end that consumers can't see, from AI chips to operating systems, to AI algorithms, AI function development tool chains, AND SaaS development platforms on the AI cloud, etc., thus providing a complete set of easy-to-use intelligent driving development chains for car companies.
And 20+ car companies fixed point cooperation, 50+ models of front installation fixed point, more than 1 million pieces of journey chip shipments, more and more users choose horizon, but also verified the correctness of the horizon mode.
Back to the chip itself, the Pilot 3 solution composed of 3 Journey 3 can already provide a very mature high-speed NOA experience with only 15 TOPS hashrate power, which makes me have to look forward to how the subsequent journey of 128 TOPS hashrate 5 and the journey of up to 1000+ TOPS hashrate will perform.
The new car equipped with the Journey 5 chip is also about to be unveiled in the fourth quarter of this year.
Horizon, we Q4 see.
-END-
(Please remember to pay attention to our video number, BMW i7 to understand)