XPeng Patents the Hard Part of Exterior Voice: Choosing Who to Serve

"Exterior voice" is the easy part. The hard part is what happens when three people are standing next to the car: the doors, the lights and the welcome sequence have to respond to someone, and picking the wrong person makes a feature feel broken rather than clever.

A XPeng patent application published on 15 September 2026 sets out how the company intends to solve it — and a Chinese technical writer, 轰鸣的小跑SVM, has written a clear reading of it. The patent, CN 122761318 A, was filed on 18 June 2026 by Guangzhou XPeng Auto Technology, and its title is deliberately dull: "a method, apparatus, device and vehicle for determining a target object."

The abstract gives away the mechanism. The car takes an image, finds the candidate objects in it, works out where each one is standing on the ground, and then chooses which to serve using a ground priority — a value that expresses how likely that person is to trigger the service. The stated payoff is selection that is low-compute, fast and accurate, which is the giveaway that this is a production problem rather than an academic one.

How the method works, in three steps

1. Find the feet, not the middle of the box. A person's position on the ground is not the centre of the detection box. The application describes looking for ankles: if both are visible, the midpoint between them is the person's ground position; if only one is, that ankle is used; if the ankles are occluded, other keypoints are projected onto the ground to estimate where the person is actually standing. It is a small detail that decides everything downstream — a person's feet are where the car's geometry has to meet them.

2. Join the image to the car. The camera sees a two-dimensional picture; the component that wants to serve — a door, a lamp, a welcome function — has a physical range on the ground beside the car. The system therefore builds the mapping between the two coordinate systems in advance, so that "the person in the frame" becomes "the person relative to this component."

3. Score the ground. Rather than measuring distance, the method lays a priority field over the ground around the car: it peaks at the mounting position of the component being served and decays outward. Someone standing where the service is meant to trigger scores high; someone visible but off to the side scores low. With several candidates, the car compares those values and takes the highest — no per-person behaviour prediction required.

The design decision worth noticing

That last point is the interesting engineering trade. A system that tried to infer what each bystander wants would be doing rich, expensive reasoning about people who are probably just walking past. This one puts the camera parameters, the mounting positions and the service preferences into the rules ahead of time, so at runtime it only has to detect objects, locate keypoints, project them onto the ground and compare numbers. The compute chain is short and — the part engineers care about — tunable.

It is also not hard-coded. If the camera moves, the component is repositioned or a service's preferences change, the mapping and the field are re-calibrated. The recognition stack does not get rebuilt because a lamp moved.

The general principle

The writer's closing point is the one that generalises beyond voice: recognition is the entry point, but spatial relations decide whether a service lands on the right person. Different functions care about different ground — parking cares about space boundaries and obstacles, a door-open warning cares about the rear flanks, low-speed assistance may rank pedestrians and cyclists highest. Narrow the field with geometry first, then spend compute where it can change what the car is about to do.

What this is, and is not

• It is a published application, not a granted patent and not a shipped feature. XPeng filed it in June 2026 and it was published on 15 September; nothing in the document says which models carry it, or whether it is deployed at all.
• The plain-language reading is a third party's. The step-by-step explanation above follows the writer's analysis of the filing; patents are drafted to be broad, and the claims matter more than the abstract.
• The document is unusually generic on purpose. The title mentions no voice, no lighting and no welcome — just "determining a target object." That is normal patent drafting, and it also means the method may be intended to cover much more than exterior voice.
• Google Patents does not serve CN122761318A yet (we checked — both language URLs return 404), so the article cites the publication number rather than a link. The two images in the original post are the application's front page and its drawings.

The bottom line

Nobody buys a car because it picks the right person to open the door for. They do notice when it opens the door for the wrong one. This is the layer of work that decides which of the two happens, and a patent application is one of the few places you can watch a car company doing it — months or years before a release note mentions anything about it.

Source: 轰鸣的小跑SVM (Weibo), explaining XPeng patent application CN 122761318 A — translated & adapted for XPENGTech.