XPeng VLA 630, Frame by Frame: Where "Fast Reactions" Actually Come From

Two everyday scenes set up the question. In the first, a car is merging but has not entered the lane yet — the XPeng G9L has already changed lanes. In the second, a delivery rider has not yet borrowed the motor lane, but the car has already started steering around him.

So is "fast reaction" about computing faster, or about predicting the world? Chinese auto-tech channel 电动知士 (Electric Zhishi) spent a week re-running typical conflict scenarios with XPeng's second-generation VLA 6.3.0 — and then reviewed the footage one frame at a time, at 4K/30 fps.

1. The latency puzzle: 267 ms, or 500 ms, or neither

In a puddle-dodging clip, an oncoming car's windscreen is barely visible behind trees. The G9L's return-to-lane trajectory appears 8 frames later — about 267 ms — and the driver's hands move on the wheel at almost the same instant. From the footage alone, the two actions look essentially synchronised.

The track test is more interesting. With a dummy falling or a push cart intruding at a 50 km/h baseline, the G9L always completed a detour or a stop — no failures, no AEB interventions. But the measured latency depends entirely on where you start the clock:

• from the moment the dummy's head appears, the trajectory change can take more than 500 ms;
• from the moment it steps onto the yellow line, the trajectory changes almost simultaneously.

On 30 fps footage you cannot even tell which came first in the second case. Measuring the same scene from a different instant turns an apparently slow system into an astonishingly fast one.

Their explanation: the system had already detected the dummy and the cart, and was continuously judging their motion trend. Stepping onto the line is confirmation that the target may keep crossing — so the trajectory changes. The gap between "head appears" and "line crossed" is not dead time; it is the model watching and updating.

2. Long temporal sequence: the past keeps informing the present

The concept the video calls 长时序 ("long temporal sequence") is simple to state: the car chains together information from consecutive moments, so what it saw earlier keeps participating in the current decision. Combined with prediction, it can judge what is likely to happen over the next few seconds and adjust as the environment changes.

The examples are all the same shape:

• A merging point. While following the car ahead, a second car pulls out from the right. Before it enters the lane, the G9L's trajectory already swings left and it takes the left lane first, pre-empting the merge negotiation.
• A cut-in. A white car cuts in ahead of the Volkswagen in front. Before it has fully entered the lane, the trajectory is already swinging left — the navigation arrow explains why: the car needs the straight-through lane at the next junction, so the left move is temporary and purposeful, and it returns to the right lane afterwards.
• A car leaving a car park on a narrow road. Before it reaches the yellow line, the trajectory has already pulled back and the car slows and stops, leaving the other driver room.
• A truck changing lane to pass a car stopped in-lane. The G9L's left indicator comes on together with the trajectory — then the indicator stops, the SR display shows "lane change cancelled", and the trajectory returns, because the left lane is narrowing.

That last case shows something worth watching closely: the indicator and the model's plan are almost perfectly in sync — on when planning to merge, off the moment the plan changes. The host suspects the turn signal may be decoded directly from the model's output, though he notes this needs confirmation from XPeng.

A single occurrence could be luck. Repeated across different roads, different targets and different types of conflict, it looks like a stable model capability rather than a coincidence.

3. Efficiency as a driving philosophy

With those foundations in place, the driving style is visibly different: fast where it can be, and behaving much like a seasoned local driver.

• Freeway merging. It began merging at the dashed line, let a fast car pass, settled into the middle lane and accelerated to the 80 km/h limit. A black VW behind had to exceed the limit to get past.
• A second merge. It signalled, spotted a Buick arriving quickly in the blind-spot view, paused the lane change, let it pass, then completed the change. It wants to go, and it can also wait.
• Jam merging. Creeping forward while watching the flow, it adjusted speed and steering with traffic, then took a gap to merge — and in a dead-stop jam it kept working its way into whichever lane was moving fastest.
• Defensive lateral positioning. Around buses pulling into stops, parked cars, wide vehicles and cyclists it offsets early whenever there is room, and it rarely out-drags cars alongside it at intersections.

Where the host disagrees: emerging from a service road at an unsignalized junction, with no traffic around, it drove straight past the stop sign without stopping (it did slow down and creep forward for visibility). He wants VLA 2.0 to learn to stop there — the one place where it has not yet matched FSD V14's stop-and-observe behaviour.

4. FSD V14, Mad Max, and "a dragon in beta, a worm at full rollout"

FSD V14 ships with four styles — Chill, Standard, Hurry and Mad Max. On US test drives the host used Hurry rather than Mad Max, to avoid tickets; V14 can be quick when the road allows and slow down when risk rises.

VLA 630's pursuit of efficiency reminds him of Mad Max — but in China's complex traffic it also shows a flexibility and negotiation ability that V14 does not. Difference in driving style comes down to training data: what road conditions and what behaviours a model has learned shape how it drives.

The video also revisits the old Chinese joke about assisted driving — "a dragon in beta testing, a worm at full rollout." When a model is not yet stable in some scenarios but has to ship to hundreds of thousands of cars, makers can tame it with post-processing and rule-based fallbacks. It is a realistic choice, and the larger the production fleet, the greater the pressure.

That is why 630's architecture change matters: long temporal sequence plus 6-second prediction lets the model handle scenes that previously needed rule-based fallbacks. The hope is that this generation keeps the flexibility and efficiency intact when it goes out to everyone.

The video closes with a thought experiment about whether Mad Max is a pretrained model's true nature or a post-training push toward the edge of its distribution — and a question for viewers: is the beta build he drove sitting at the model's behavioural centre, or out in its high-efficiency tail?

Source: 电动知士 (Electric Zhishi), Weibo — translated & adapted for XPENGTech.