Chapter 243: LAD (Luna Auto Drive)

Park Cheol-jin manipulated the meeting room PC to display the data.

A heatmap of driving trajectories appeared on the screen, dotted with points across the Korean Peninsula.

Thick lines of highways and dense dots of downtown areas squirmed like living blood vessels.

“I’ll start with the current status.”

Park Cheol-jin flipped through the pages, explaining the sales status of the EV-Nova vehicles.

“Currently, the cumulative sales of EV-Nova are 21,204 units, with 3,318 units shipped last month, and this month is maintaining a pace of 3,000 to 3,500 units.”

Every time I heard this, I couldn’t hide my disappointment with the production capacity.

Demand was overflowing, but they couldn’t sell more due to lack of supply.

However, it would still take another 4-5 months to complete the capacity expansion.

“How many of the delivered vehicles have agreed to ‘LAD (Luna Auto Drive)?’”

“18,742 units, and the monthly active units (MAU) are 12,406.”

Park Cheol-jin pressed the remote control.

Detailed data appeared.

[Cumulative Driving: 890 million km]

• LAD (Luna Auto Drive): 210 million km

• Rare Events Mined: 78,412

After slowly reviewing the data,

I pointed to the part of the heatmap where the red dots were clustered.

“There’s a lot of driving data, but the important thing is quality over quantity. We need to focus on accurately selecting ‘moments worth learning’ from this.”

Kim Se-bin immediately continued my words.

“For that purpose, we’re constantly running ‘Shadow Mode’ on all vehicles equipped with LAD. We compare the steering and acceleration/deceleration made by the actual driver with the route suggested by Luna on a frame-by-frame basis and extract only the sections where the discrepancy exceeds a certain threshold.”

Kim Se-bin flipped the slide.

A new graph appeared.

[Shadow Mode Discrepancy Rate (Last 30 Days)]

• Highways: 22.3% → 8.5%

• Downtown: 38.2% → 15.3%

I glanced at the graph and nodded.

“The discrepancy rate is higher than I thought.”

It was somewhat expected since it was still early.

That’s why we needed to first confirm why the numbers were so high.

Kim Se-bin added further explanation.

“The main reasons for the high discrepancy rate are summarized on the next slide.”

The next slide categorized the main reasons into three.

1. Unique intersection rules of domestic roads

• Many unprotected or signal-less left turns, narrow U-turn pockets, and variable lanes, with signal positions and lane markings varying by region

2. Irregular driving of two-wheelers and PM (Personal Mobility)

• Many patterns outside the norm, such as driving in reverse, passing between lanes, and encroaching on sidewalks, and tracking is difficult at night due to differences in headlight brightness

3. Temporary lanes and barriers in construction zones

• Paint and reflective tape are mixed, and on rainy days, road surface film or glare makes it difficult to distinguish lanes

I silently read the data compiled by Director Kim Se-bin in detail.

I knew all of these issues.

Even before the regression, I had seen them in papers and numerous documents.

I got up and walked over to the whiteboard.

“Solving the discrepancy rate problem isn’t just about collecting a lot of learning data.”

I slowly continued, drawing four circles on the whiteboard.

“This needs to move in four directions simultaneously: data, sensors, driving planning, and infrastructure. These four axes must be balanced for autonomous driving to work properly.”

Director Kim Se-bin and Park Cheol-jin stared at the whiteboard, trying to understand my words.

I added further explanation.

“As I said earlier, what we need isn’t a lot of data, but precisely selecting the necessary data.”

I turned my head to look at Park Cheol-jin.

“You said rare events mined were 78,000 cases, right?”

“…Yes.”

“From those, focus on extracting three situations: unprotected left turns in downtown areas, sudden entry of two-wheelers, and lane recognition failure in construction zones. And this shouldn’t just use actual situations, but should be expanded with digital twins.”

Han Yu-ju, who arrived late to the meeting, asked in mild surprise.

“Digital twins, do you mean a virtual space?”

“Yes. For example, even for the same intersection, just by changing conditions like time of day, weather, glare, road surface film, and backlight, we can run simulations in dozens or hundreds of versions. That way, we can secure many more rare events than now.”

If the number of sold vehicles were in the millions, we wouldn’t need to do this, but

with fewer than 100,000 vehicles, this method was the best we could do.

“But… wouldn’t using synthetic data reduce learning accuracy?”

As expected of Dojin Soft’s leader, the question was sharp.

“That’s why we also need to change the labeling method. Previously, a person would watch the entire video and mark ‘this is dangerous,’ but from now on, if Luna first says, ‘this section looks dangerous,’ a person will only confirm that part.”

“Ah… you mean changing to a method where a person verifies Luna’s hypothesis. That would definitely be more efficient.”

Luna first tags numerous drives with ‘this seems odd,’

and a person only checks if it’s right or wrong.

I named this method the Hybrid Labeling Structure (HIL).

Then, I pointed to the second circle.

“The second is sensors. Rather than changing the hardware, what’s urgently needed now is alignment.”

“Alignment…?”

Park Cheol-jin tilted his head.

I drew a simple chassis diagram on the whiteboard along with an example of misaligned camera angles.

“If you carefully examine the shipped vehicles, you’ll find that the camera angles are slightly misaligned by about 0.5 degrees.”

Within the standard level that causes no problems for general auxiliary functions,

but it could be fatal for autonomous driving.

‘Even before the regression, it took Tesla more than two years just to figure out this was the problem.’

“Are you referring to BEV projection errors?”

Kim Se-bin, being an expert, was the first to catch what I was trying to say.

“Yes. Even if the camera angles are misaligned by just 0.5 degrees, the projected lane positions are displayed differently from reality. That causes errors in steering commands, and the more it repeats, the higher Luna’s discrepancy rate becomes.”

I drew another small chart on the whiteboard.

“And the bigger problem is that the errors accumulate. For example, if Luna keeps suggesting odd routes on road A, it’s not a map or model issue but that the sensor’s reference coordinate system itself is misaligned.”

“Then… do we need to announce a recall to adjust the cameras one by one?”

Park Cheol-jin looked at me with concern.

I shook my head.

“That’s the last resort.”

I picked up a board marker, drew a square on the whiteboard representing the ‘alignment correction algorithm,’ and wrote ‘remote alignment’ below it.

“This needs to be done via OTA (Over-The-Air). During nighttime parking, when the vehicle is stationary for a certain period, activate the routine. Project a specific light pattern onto the front of the vehicle, and have the camera recognize it to reverse-calculate the distorted angle of view.”

“A pattern projector will be needed.”

At Park Cheol-jin’s words, I nodded.

Kim Se-bin added an explanation.

“Actually, we’re already testing that in a pilot. I was also guessing that might be the problem.”

“Huh… really?”

An unexpected answer.

I let out a low gasp.

“Then, based on the data collected from the pilot vehicles, is BEV mapping being corrected in real-time?”

Park Cheol-jin asked with interest.

“Precisely, Luna recognizes and reflects the correction offset. That’s the core of alignment.”

I emphasized by drawing a thick border around the second circle with my finger.

“No matter how good the data and algorithms we have, if the sensors are distorted, it’s all useless.”

If your eyes are misaligned, it’s only natural that the world looks strange.

There was a brief silence in the meeting room,

and I immediately moved on to the third solution, driving planning (Planning).

I pointed to the third circle drawn on the whiteboard.

“More than anything, the most important thing in autonomous driving is ‘responding to uncertainty.’ Especially in downtown areas, and unpredictable objects. For example, two-wheelers like bicycles, kickboards, children, and construction workers.”

I briefly manipulated the computer to display the ‘sudden pattern heatmap’ on the screen.

“If the goal of autonomous driving development so far has been ‘not to collide,’ now it needs to change to ‘avoid first.’”

“Do you mean strengthening the defensive driving concept?”

At Kim Se-bin’s question, I nodded.

“Especially, introduce a history-based defensive weight. Even if it’s the same PM (Personal Mobility), an object that has suddenly appeared before is likely to be dangerous again, so Luna remembers that history and reacts earlier next time.”

It might seem like a simple act,

but it’s a really important part, and one that’s easy to overlook.

“Adding past behavior history to the deep learning model…”

Kim Se-bin’s expression noticeably brightened.

It was probably because he felt that autonomous driving was evolving to remember and avoid like a human.

I paused for a moment, then pointed to the fourth and final circle on the whiteboard.

“The last is infrastructure. This can’t be directly laid by Dojin Frontier, but we can receive the data.”

Next to the fourth circle, I added the words ‘external DB linkage for dangerous sections.’

“By linking with traffic centers, road construction companies, and even delivery apps and navigation companies, LAD (Luna Auto Drive) vehicles will know in advance where the dangerous sections are before they start driving.”

Han Yu-ju immediately responded.

“We’ll support that part.”

“Yes, with Dojin Soft’s support, it can be resolved within a week.”

Dojin Soft had already dominated the app ecosystem based on the smartphone OS ‘Lunar,’ including maps, delivery, shopping malls, taxi calls, weather, and community platforms.

By linking that vast API and user location data with the LAD system,

we could dramatically improve real-time danger prediction accuracy.

After finishing the explanation, I slowly scanned the four circles in front of the whiteboard.

[Data - Sensors - Driving Planning - Infrastructure]

Four core axes that would collapse if even one were missing.

Still early, but the framework was gradually taking shape.

I turned my head to face everyone and said,

“For the next month, after reflecting the solutions I mentioned, we’ll conduct a public autonomous driving test.”