Over the past two years, discussions on autonomous driving technology have inevitably revolved around a few key terms: end-to-end, VLA, world models, and increasingly expensive training GPUs.
The discussions have focused almost entirely on the model side.
Yet, few have directly asked a crucial question: If the model changes, do the sensors need to change accordingly?
Recently, news emerged that Afari Technology has launched an AI millimeter-wave radar (4D). Afari claims this is the first time a 10 TOPS NPU has been integrated into a radar.
I read through all the available materials from start to finish and cross-referenced Afari Technology's acquisition announcements and partnership information. After reviewing everything, I believe this radar serves as a perfect case study for this very question.
Therefore, this article delves into how the data sensors feed into models should evolve in the post-end-to-end era.
Summary in one sentence
Current approach: Sensors process data according to their own standards, leaving the model to adapt.
Afari's approach: Moves the decision-making step into the radar, delivering the 4D features required by the model.
Raw data is not lost either: The satellite mode sends intermediate frequency data directly to the central domain controller.
Let's break it down and analyze.
1. Incorrect Formats Force the Model to Pay the Price for Sensors: Fine in the Rule-Based Era, but the Cost Changes Post-End-to-End
Let's first look at the current common practice.
Sensors first process data according to their own standards: cameras output images, while millimeter-wave radars and LiDARs output point clouds. The model then unifies these formats in either early fusion or late fusion.
In the rule-based era, this workflow had no issues. The decision-making logic was written by humans, so engineers could simply write code based on the sensor outputs.
End-to-end is different. It feeds camera pixels directly into the model, eliminating the intermediate pipeline of "detect bounding boxes first, then track, then plan." The output is directly the trajectory.
With decision-making handed over to the model, problems arise: The information discarded by sensors at the front end according to fixed rules must be compensated for by the model at the cost of computing power and accuracy, and some of it simply cannot be recovered.
To use an analogy, it's like a chef taking over a batch of pre-prepared ingredients, but if the ingredients are wrong or the proportions are off, no amount of culinary skill can produce an authentic dish.
Simply put, in the post-end-to-end era, what the sensors feed into the model begins to determine the upper limit of the model's performance.
2. Vision Is So Powerful, Why Is Radar Still Needed? What It Measures, Cameras Cannot
Cameras rely on passive light sensing, capturing 2D projections along with brightness.
Distance must be inferred through triangulation or priors about object dimensions, and speed must be deduced from the displacement between adjacent frames. This is sufficient in good lighting, but in rain, fog, backlighting, or at night, once the signal-to-noise ratio drops, errors increase.
Millimeter-wave radars precisely compensate for this. The electromagnetic waves at 77GHz have a wavelength of about 4 millimeters, offering far superior penetration through rain and fog compared to visible light.
More importantly, it relies on the Doppler effect to directly measure speed. What is measured is a physical quantity, with no inference errors.
There is a huge difference in how to handle a lead vehicle that is stationary versus one that is creeping slowly, and Automatic Emergency Braking (AEB) is most vulnerable to errors in this exact judgment.
Does the radar itself have flaws? Yes.
Traditional 3D radars only measure distance, speed, and horizontal azimuth, lacking elevation angles. They cannot distinguish whether the echo ahead is an iron plate on the road or a person standing on it. The angular resolution is also coarse, causing multiple targets to blur together when they are close to each other.
The result is two types of misjudgments: treating iron plates, manhole covers, and gantries as obstacles, leading to "phantom braking"; or merging side-by-side cars and two-wheelers into a single target, missing the dangerous one.
The solution for many automakers is to downgrade the radar's weight, giving it less say in AEB.
4D radars add the elevation angle, giving radars height information for the first time. Their installation volume has been climbing rapidly in recent years, although the statistical calibers of various institutions vary widely, so I will only discuss the trend here.
The radar remains on the vehicle relying on physics; its problem has always been that the data it feeds to the model is too coarse.
3. Which Link Does Afari's 4D Radar Change? Letting the Model Decide Which Echoes Are Worth Keeping
Afari claims that for this radar, the entire pipeline from signal detection to target classification is handed over to the model.
This statement needs to be broken down, because many "AI radars" on the market merely add a model layer after the output of the traditional pipeline for secondary judgment, while the front end still uses the old methods.
01 Where the Traditional Pipeline Gets Stuck
Traditional radar signal processing involves five steps: ADC sampling, FFT transformation, CFAR threshold detection, clustering, and tracking.
The core is CFAR. It determines which echoes count as targets based on a fixed threshold.
If the threshold is set too high, weak targets such as pedestrians, two-wheelers, and low-reflectivity cones will be filtered out; if set too low, clutter and multipath ghost images will all be retained. Parameter tuning can only pick a position between these two extremes.
An iron plate is a typical example: the iron plate has strong reflection and steadily exceeds the threshold; the person standing on it has a weak echo and might instead be suppressed.
02 After Switching to a Model
The basis for judgment changes from "whether the echo intensity exceeds the threshold" to "whether the features of this group of echoes resemble a target."
The model can learn during training what multipath ghost images look like and what patterns the echoes of weak targets have, no longer relying on a single number for a one-size-fits-all cut.
Afari claims that this radar can distinguish between an iron plate and a pedestrian standing on it; for small targets like children, where traditional 3D radars can stably detect at about 50 meters, it can achieve 100 to 150 meters.
The deliverables have also changed accordingly. The 10 TOPS NPU completes point cloud parsing, target classification, trajectory prediction, and occupancy estimation at the radar end. The output changes from raw point clouds to 4D features aligned with the model.
I believe this is the most critical change in the entire radar, allowing the model to decide which echoes are worth keeping.
4. Why Not Hand Over All Raw Data to the Model? Trainable Processing and Hard-Coded Rules Are Two Different Things
Reading this far, industry insiders might ask a question.
Isn't end-to-end about letting the model learn on its own? Then why not just hand over all the rawest echoes directly to the central large model? Does processing it first at the radar end not constitute another form of information loss?
My understanding is: The difference lies in whether this step can be trained.
CFAR is a hard-coded rule, and what is discarded is predetermined by engineers. The model at the radar end can be trained and upgraded together with downstream models, with what to discard and what to keep determined by the data.
Additionally, Afari has not blocked the path of raw data; it provides two modes.
• Edge computing mode: The NPU completes the judgment at the radar end and directly outputs structured results. The pipeline is short, the response is fast, and there is no need to transfer massive point clouds.
• Satellite architecture mode: The radar only collects and transmits, sending intermediate frequency data without target detection to the central domain controller for unified processing. The information is more complete, and the algorithms can be unified.
Both modes use the same set of hardware and are selected based on the vehicle's computing power configuration. Furthermore, in satellite mode, the computing power at the radar end remains online and can issue its own warnings when the domain controller or the link is abnormal, although the main path remains central fusion.
OTA is the same. Once detection is handed over to the network, upgrades modify the model parameters. I infer that upgrades will be more granular and faster, but this must be verified by mass-produced vehicles.
What Afari answers is what data to feed; how far it can see is merely the result.
5. For This Answer to Hold, What Else Is Needed? The System Can Run, but Three Proofs Are Still Missing
Let's first look at what is already in place.
• Algorithm: The foundation model of StepFun. Both parties announced in April 2026 the co-development of a native autonomous driving foundation model.
• Perception: On June 11 this year, Afari Intelligent Driving acquired 100% equity of Ronggan Technology for CNY 25.908 million.
• Vehicle: Scenarios and data from the Geely system.
Afari handles the autonomous driving algorithms, edge-side chip adaptation, and system integration itself. What data the model needs, how to modify the radar, and which vehicle to use for verification after modification can all be circulated within a single system.
The path it takes lies between full-stack in-house development and pure integration. It uses equity and a unified development system to link several companies together, grasping the right of definition itself. The advantage is not having to invest from scratch, while the cost is high coordination expenses.
The system is already operational; we will wait and see the three proofs of mass production, cost, and external customers.
6. In Conclusion: The Answer Lies in the Installation Volume
End-to-end has pushed the model to the leading role in autonomous driving. With the lead changed, what the supporting roles deliver must also change.
What sensors used to do was to hand over what they saw as completely as possible; what needs to be considered now is delivering data in the format the model requires.
Afari's radar is a hardware sample of this transition. Of course, the parameters are written in the release materials, and on-road performance still needs to be proven by mass-produced vehicles.
However, this time Afari has also opened up a new product and technology direction as a rising star in China's autonomous driving.
That's all for today's sharing. If you find it useful, please help by liking, sharing, and clicking "Wow", and you are also welcome to discuss your views in the comment section.
* Reproduction and excerpting are strictly prohibited without permission.