Folks, can you imagine how affordable smart driving has become?
Not long ago, we were saying that autonomous driving was just a toy for the wealthy.
Car manufacturers would casually throw in a couple of lidars, costing tens of thousands of dollars in actual expenses. Even if car owners opted for them, starting at tens of thousands of dollars, they weren't guaranteed to be effective.
Not long after, the price of smart driving hardware plummeted, becoming as cheap as cabbage, just like my fund.
Recently, DJI Drive released several new products on their "Ready to Go" platform. One of them, a pure visual version using only cameras, not only supports advanced features like highway and urban NOA, but also doesn't rely on high-definition maps and consumes very little power. Basically, you can slap it on almost any car and use it.
The kicker is, the price is less than 7000 CNY, about the cost of the air springs found in many cars.
If I were an automaker, I'd definitely go for advanced intelligent driving over just raising and lowering. But here's the catch: they say you get what you pay for. Take Tesla's FSD, which Elon Musk has been refining for years and still sells for a hefty sum. On the other hand, DJI, a newcomer in mass-producing smart driving tech, not only offers cheaper hardware but practically gives away the software.
How come?
Just as I was pondering this, DJI goes ahead and sends out test drive invitations. Can I really resist that?
After all, most of the smart driving systems I've tried before were, well, less than impressive. Many tend to act up now and then, making nearby lanes jittery and demanding constant attention.
DJI's purely visual solution didn't even bother with depth validation by LiDAR, so it was ripe for criticism due to the likelihood of bugs galore. But DJI silenced the critics with just over 30 minutes of test driving.
The test vehicle was essentially the Baoyun Cloud's intuitive driving version, with the key difference being the replacement of the original car's 32 TOPS low-power domain controller with a high-power controller featuring the Qualcomm 8650 chip, boasting 100 TOPS.
Compared to the mainstream solution featuring a dozen cameras, one or two LiDARs, and N NVIDIA chips boasting 254 TOPS, this hardware setup might seem a bit modest.
However, in terms of user experience, it's one of the most user-friendly autonomous driving systems I've ever tried.
The test drive route took me on a loop through the bustling streets of Shenzhen. The journey lasted around 30 minutes, passing through numerous intersections, including some peak congestion zones where traffic came to a standstill.
However, throughout the whole journey, I never once took control.
On relatively simple straight roads, its performance is actually not much different from other models. It accelerates and decelerates smoothly, without sudden nods or pushes, and changing lanes doesn't give passengers any noticeable tug.
However, it's in the details where it really shines.
For example, when encountering a red light, it first slows the car down to a very low speed, like below 10kph, then gently maneuvers to a position close to the preceding car before coming to a complete stop.
I asked the engineer riding along about it, and he said their algorithm calibrates when the camera can just no longer see the rear wheels of the car in front, which is roughly a distance of 3 meters.
He seemed genuinely pleased, as that's how I usually drive.
Another thing is yielding to larger vehicles. During the test drive, we passed through several stretches with heavy truck traffic. Each time a large truck passed or we needed to overtake one, the car would slightly veer to avoid it.
Moreover, the degree of this avoidance isn't fixed. If the relative speed is high, the maneuver is subtle, almost imperceptible. But if the relative speed is low, the car noticeably swerves, evoking a sense of social anxiety, like wanting to steer well clear.
The algorithm engineers at the self-driving platform probably haven't missed the highlight reels of big car accidents. But compared to these minor details, I think its real godlike ability lies in its super-fast gaming speed. No fuss, just decisiveness.
Friends who often drive know that the trickiest part of making a left turn at a major intersection is not knowing which lane the cars next to you will turn into. If the car on the left turns wide and the one on the right turns tight, it's like being sandwiched between two slices of bread with cheese. You have to make frequent micro-adjustments.
What's amazing about this DJI intelligent driving system is that it basically replicates all the maneuvers I would do. When the car on the left gets too close, it slightly veers to the right. After getting in position, it even adjusts to the left again based on the position of the car on the right.
When I glance at the left rearview mirror, it deals with the left-side traffic, and when I check the right mirror, it starts handling the right-side cars.
It's like the vehicle and I are in sync.
Encountering pedestrians crossing the road is similar. Other autonomous vehicles seem to possess impeccable manners, always waiting patiently for pedestrians to cross first.
Once the pedestrian crosses, it's fine. But if both sides intend to yield, there's a deadlock, with neither side moving.
DJI's approach, however, is more assertive: "You yield to me? Then I'll proceed."
In the few pedestrian encounters I've had, if a pedestrian yields to the vehicle, the test drive cars accelerate decisively through. While it may not seem as courteous, it undeniably boosts traffic efficiency.
Then there's what I think is the coolest part: evasion from behind.
During the test drive, there was a scenario like this: driving straight on a lane, and suddenly a small electric scooter darted across from the left.
Normally, you'd just swerve to the right, right? But at that moment, there was also a bus behind preparing to accelerate past. Changing lanes would definitely lead to a collision.
Typically, regular automated driving systems would throw in the towel, requiring human intervention to stop or change course. But DJI's decision was to halt the lane-change process, brake to a safe speed, wait for the bus to pass, and then resume driving.
There was another scene where I almost got into an anime-like situation.
We needed to make a right turn into the far right lane. My focus was on observing the behavior of multiple lanes merging, completely oblivious to the fact that there was a mud truck on the side also intending to take this same right turn.
Luckily, the rearview camera came to the rescue, and the smart driving system swiftly steered the car back, making room for the mud truck with just a narrow gap.
The onboard engineer praised me for being bold and meticulous, but honestly, I just didn't see it.
After driving around the track, I feel that overall, this system is detailed, operational, and remarkably stable. If the final version pushed to production matches the performance of this test drive, I might actually use it frequently in my daily driving.
As for why DJI can achieve advanced autonomous driving with just a few cameras and minimal computing power, they briefly touched upon it during the press conference. Something about optimized algorithms and high perceptual accuracy, which sounds quite mysterious.
One intriguing aspect is a technology called neural network trajectory prediction. It involves a Transformer model embedded in the high-computing power domain controller, capable of predicting the motion trajectory of objects based on what the camera sees, using AI.
Then it assesses potential conflicts between the future trajectory and the driving path, and adjusts the current vehicle accordingly.
It's quite impressive, isn't it?
But rather than delving into how they achieved this, I think the more important message conveyed by this test drive is that the previous approach of car manufacturers relying heavily on hardware and computing power for autonomous driving might have been off the mark.
For a long time, car manufacturers have been plagued by a serious "firepower deficiency phobia" when it comes to autonomous driving. In order to gather as much information from the road surface as possible, all kinds of cameras and LiDAR sensors have to be maxed out. Take Avita, for example; it has three LiDARs on a single vehicle alone.
To tackle various challenging problems, a high computational power chip must be installed to accommodate a sufficient number of rule algorithms. For instance, in the case of NIO, Nvidia's OrinX with 256 TOPS computational power requires four chips just for this purpose.
However, in many vehicle models, a significant portion of computational power is allocated not to "solving problems" but to "understanding what the problem is."
Here's an example to illustrate:
In vehicles equipped with various sensors such as LiDAR, cameras, and other hardware, there is a process of analyzing the collected information collectively. This process involves aligning the timeline, ensuring spatial consistency, and determining whose information takes precedence.
For instance, the accuracy of lateral motion recognition by cameras surpasses that of various types of radar. However, in adverse weather conditions like dust or rainy days, LiDAR performs better due to its superior visibility.
In different scenarios, sensors need to align their granularity accordingly.
This extends beyond just coordinating between different sensors; even within the camera system, there are numerous coordination issues to resolve. For instance, identifying targets in the same lane relies on the forward-facing camera, while recognizing targets in adjacent lanes requires the side-view camera.
When the distance is too close, radar encounters echo problems, necessitating the use of parking cameras.
In essence, it's all about multi-sensor algorithms. The more scenarios engineers can think of, the more complex the rules for hardware coordination become.
Similar to circuit design, ensuring a rule runs smoothly requires designing a plethora of validation functions to guarantee its implementation.
For automakers lacking strong algorithmic capabilities, it often ends up as a code mountain, resulting in a significant waste of computing power.
The algorithmic prowess has finally come to its senses lately. Building a towering heap of dung isn't exactly a solution.
Instead of cluttering things up with a bunch of sensors that neither excel nor suffice, it's better to get down to brass tacks and optimize from the ground up, maxing out the capabilities of a select few sensors.
This approach not only eliminates the need for complex coordination but also reduces the demand for high computational power in chips and saves a fair bit of cash that would otherwise be spent on radar and cameras.
That's why companies like Huawei and DJI are now leaning towards relying more on higher-precision visual data, cutting down on the number of lidars or even ditching them altogether.
And, lower computing power and hardware requirements also mean that more affordable models can now be equipped with autonomous driving capabilities.
For instance, Volkswagen has announced that their Tiguan will be equipped with DJI's autonomous driving technology by 2025. A friend of mine within the Volkswagen Group also discreetly mentioned to me that Porsche and Audi are currently considering adopting DJI or Huawei's autonomous driving solutions.
Perhaps soon, with just tens of thousands of CNY, BYD or Wuling Hongguang could also offer advanced intelligent driving capabilities.
With that in mind, if there are still any new cars boasting about their extensive use of lidar, chances are their algorithmic prowess might not be up to par, relying more on quantity than quality.
Who's being talked about? Brothers, feel free to recognize yourselves.
