Unexpectedly, today's Xiaomi conference directly released the phased results of automatic driving; The overall progress is faster than I thought, which must be well analyzed. (Please watch this article with the original video, the effect is better)
Let's start with the conclusion:
1. In terms of automatic driving technology, Xiaomi's current state of display has no advanced breakthrough in the industry.
2. The overall level is in the upper class of ROBO Taxi modified vehicles.
2, but considering the time of the test car, the surge in progress is still very strong.
Modified car research
In the video, there are two kinds of Xiaomi self-driving test vehicles, and the sensor arrangement form is relatively primitive - a "big hat" is installed on the roof of the car, and the lidar and camera are installed on it.
This is the usual type of modification used by the ROBO Taxi type "autonomous driving company", and it is also more suitable for Xiaomi, which has no prototype for the time being.
The advantage of this is to give the sensor the maximum perceptual space; The disadvantage is that the family car does not allow this manufacturing mode, and the sensor must be integrated into the industrial design of the whole vehicle, so this state to the whole vehicle also needs a long period of engineering integration and cumbersome testing.
To put it bluntly: we cannot currently guess the industrial design, sensor layout and other information of Xiaomi's 24-year new car from the state of the self-driving test car.
CNOA status research
The first paragraph of the video is the auxiliary state of the urban pilot, the so-called CNOA (City Navigate on Autopilot); This is the state that everyone wants most: open the car map navigation, select the destination, and the car will drive fully automatically.
Because the progress is the first, Xiaomi's CNOA must rely heavily on high-precision maps; Pay attention to the visualization in the lower right corner, the 3D point cloud of the road surface is unusually stable, it is unlikely to be real-time modeling, and it is speculated to be a visual reconstruction of the HIGH-definition map.
But Xiaomi's perception part is really well written, please note that the visualization here, the recognition reconstruction of other traffic flows, there is no perceptible violent shaking.
It shows that the perception system of the vehicle is very stable, can accurately grasp the target, and the visual smoothing algorithm of 3D reconstruction is in place; And the model is more elegant, and the visual inspection of fluency is also very good, I like it.
Also pay attention to the three-dimensional frame set on the 3D model, which is also the basic exercise; Previously, the autopilot demonstration video of a car factory was a two-dimensional frame for target recognition, and the industry peers laughed at it.
Avoid large cars and jams Cut in, routine operation, Xiaomi test car perception conditions far beyond family cars, can do not be unusual, can not do is the problem.
NoA's signature operation – self-decision-making on-ramp and active lane change.
Xiaomi's regulation here is doing a good job, and the lane change strategy is not to look at the distance in front of the car rigidly, but to look at the relative speed; Although the left lane has not yet exceeded its own lane, the system observes that it is relatively fast, and immediately changes lanes to improve traffic efficiency.
This strategy is still relatively clever, but the video acceleration is too fast to see how smooth the lane change control is.
Risk avoidance, normal operation; Everyone is desperate to upgrade the perception system, and make lidar + surround view + millimeter wave multi-layer perception, which is what does this.
Unprotected U-turn, this is a high-difficulty scenario, you need to perceive the system to work at full strength, while the decision-making system should take into account the efficiency of traffic and safety - can not turn too slowly, otherwise you will not be able to fall in this life; You can't turn too fast, or something will happen.
Thanks to the extremely superior observation position on the roof, the test car's perception system provides 360-degree information, and this U-turn is relatively smooth.
But if installed on Xiaomi's mass production car, Lidar can only provide observation information in the front hemisphere at most, and millimeter wave also has a bunch of dead ends, which has to rely purely on surround vision perception; This is what I said earlier, the engineering problems that Xiaomi needs to solve on the production car.
There is a visualization of traffic lights (lower right corner), but it is not known whether it is the information sent back by V2X (running in the demonstration section of Wuhan autonomous driving) or the information based on pure visual recognition.
If it is the latter, then xiaomi's technical state is still OK.
Avoid the bus (big car), Ki
This paragraph is very interesting, pay attention to the visualization information, Xiaomi test car has established a perfect pedestrian, two-wheeler identification model. Can recognize, can avoid, can 3D reconstruction, the progress is good.
Complex road conditions turn left, but there are traffic lights, the difficulty is medium.
This paragraph is very interesting: thanks to the well-located perception system, the Xiaomi test car was the first to find an electric vehicle that was obscured by an SUV.
This recognition is very fast, and the decision to continue to drive forward is also very decisive, I don't know if the production car can reach such a stunning level.
Pedestrian avoidance, basic exercise.
Roundabout driving, CNOA's base exercise.
I'm heading toward the road with my own national conditions here! Super busy road section, mixed people and cars, Xiaomi test car performance is very stable, regulation and control is doing well.
Turning right without protection, I initially thought there was a common bug - that is, the big brother who recognized the red circle, did not stop in time, and drove to the front of the person before stopping.
Later it was found that it was not, it was detected that the left front of the pink electric car, must be avoided, the safety of the prejudgment algorithm written really does not poke.
CNOA status conclusion
The CNOA of the Xiaomi test car looks relatively smooth overall, which is a typical urban high-precision map strong dependence route, and because the sensors are all placed on the roof, the perception conditions are very superior.
There are many manufacturers that can achieve this level, and ROBO Taxi (the so-called "L4 self-driving startups") has grabbed a lot of them.
Xiaomi's better point is that they are very tight on time, and they take out this set of things within 500 days.
The perception, decision-making, and control algorithms seem to have a fair maturity, and it is worth affirming that they can run to this extent.
The challenge of Xiaomi is that in the process of engineering mass production, sensors, especially lidar, must be highly integrated with the design of the vehicle, so the perception space will be greatly limited.
Lidar will change from the current 360 degrees without dead angles, to only see the front hemisphere at most 270 degrees (1 front 2 side layout), or even less than 180 degrees (roof forward single Lidar layout); Therefore, the perception algorithm of surround vision puts forward very strict requirements, and it is expected that Xiaomi mass production vehicles will give a satisfactory answer.
AVP automatic parking judgment
The second part is AVP unmanned parking, that is, as shown in the picture - standing at the garage door and clicking the mobile phone, the car automatically parks into the reserved parking space.
Because the garage environment is relatively simple and the speed is low, the test car simply drives forward in an unmanned form.
Automatic descent downhill I still think it is quite amazing, the perception system is so strong?
However, looking at the second picture, emmm, is still pre-built - that is, there is already a high-precision map of the garage in the car, which is much less difficult.
This picture will be a little more obvious, the vehicle is still completely downhill, there is still occlusion on the left, and the visualization of the garage has come out (bottom right corner).
However, the perception system is OK, and the speed of recognizing the vehicle is quite fast.
This one will be more obvious: the outbound vehicle will be recognized when it crosses the line, and there are two empty parking spaces that are also accurately identified; Even if there is a high-precision map blessing, the perception system is not good, it is useless.
90 ° sharp bend, the steering wheel pulled out a huge angle, the control algorithm is very powerful; But after all, the test car, the situation of the mass production car can not be inferred from the video.
Circumvent the ice cream tube, AVP base exercise
Reverse into the garage, basic exercise
The mechanical arm automatically charges, thinking very well, and the practice is very difficult: the position of the car's charging port is not uniform, there are distributions between the front and back and the left and right, and the algorithm of the robot running around alone is enough for you to drink a pot.
And the charging port cover of many cars is not electric ejection, how do you open it? Even if it can be ejected electrically, how do you activate the pop-up program?
Not to mention that the national standard charging port standard is written very ha batch, the level is low, and the locking mechanism is abnormal, resulting in the charging gun locking, which is all basic exercises; In case of locking, the owner has to go to great pains to unlock it urgently, and even can't solve it...
In short, if you want to use the form of a robotic arm to do third-party public automatic charging, it is not very realistic, and in the long run, this is still a very heavy manual work.
However, if it is rich B to install their own home, then when I did not say, at this stage, plugging in and unplugging the charging gun is indeed not elegant enough... It would be much better to have robots do the work.
AVP status conclusion
Compared to cnoa status, the AVP status of automatic parking has fewer surprises; The main thing is that you built the map in advance, and did not see that the real-time perception algorithm has too much room to play.
Final conclusion
The 8-minute video displayed by Xiaomi shows the status of the engineering modification test vehicle in its entirety, and also covers the three scene fragments with the largest number of assisted driving at present - "urban area - high-speed (ring road) - garage".
As I said before, it is very difficult to infer a production car from this state; Xiaomi's biggest purpose should be to build public confidence and express its willingness and ability to stand in the first echelon of high-level assisted driving.
The capabilities shown so far do not have strong forward-looking breakthroughs; But again – do you think about time?
And people are not here, as a new car factory, the test car is the most important thing.
From the official start of the team to show the status of the test car today, I think the xiaomi car ADAS progress is quite fast, faster than I expected; I thought it would start showing in the first half of 23 years, and the camouflage car test would start in the second half of 24 years.
Eye measurement production vehicle is still the first echelon of perception hardware: the front hemisphere Lidar + surround vision + millimeter wave of the 2-layer semi-architecture; I'm more interested in which computing platform to use, when the old yellow monolithic 1000T hashrate Atlan should come out, but Horizon J6 and Qualcomm's next-generation Ride will also appear.
The others don't guess, the information is too little, let's go here first.