Leapmotor logged 1 million km of driving data, but after filtering and labeling, only about 40,000 km remained as reusable scenario knowledge. The company is turning data gathered around the world into validation assets that can be applied across markets, vehicle programs, and future projects. As its business with Stellantis expands, Leapmotor's ability to enter new markets increasingly depends on reusing road experience across markets, vehicle programs, and future projects — not simply deploying more test vehicles.

By Sang Min Han _ han@autoelectronics.co.kr
한글로보기






Leapmotor moved quickly into overseas markets. It signed a global strategic partnership with Stellantis in October 2023, and by February 2024 the two companies were jointly tuning chassis technology. Leapmotor International launched that May, and the first batch of Leapmotor EVs was shipped from China to Europe in July. The C10 and T03 entered 13 European countries; the B10 had its world premiere at the Paris Motor Show. Sales later expanded into South America and Mexico.
But shipping a vehicle overseas is not the same as ensuring consistent performance across global markets. Regulations, road conditions, climate, traffic culture, and driving habits all differ by country. A vehicle thoroughly validated on China's complex roads can't be assumed to handle Europe's roundabouts and tram tracks, the Middle East's heat and sand, Southeast Asia's monsoon flooding, or long-distance highway driving in the Americas without further proof. Shi Wei, Leapmotor's head of global validation and R&D digitalization, turned that problem from a matter of test records into a matter of assets.
"Data is not an asset."



Leapmotor said that only about 4 percent of its 1 million km of raw driving data — roughly 40,000 km — remains as reusable scenario knowledge after filtering and labeling.   



The 4% that remains from 1 million km

Leapmotor said it has built a global scenario asset system spanning 32 countries, including China, and 195 cities. Cumulative driving distance grew from 100,000 km in 2022 to 300,000 km in 2023, 600,000 km in 2024, and 850,000 km in 2025, reaching 1 million km in 2026.
But Shi didn't count all 1 million km as an asset. Raw data collected on the road only becomes an asset, he said, once it has been filtered, refined, and labeled to the point where it can actually be reused in development and validation.
The filtering process Leapmotor described is unforgiving. Treat raw data as 100 percent, and only 35 percent remains as meaningful scenarios. After labeling for validation use, the share falls to 12 percent, while reusable scenario knowledge accounts for just 4 percent. This leaves roughly 40,000 km of usable validation assets from 1 million km of driving data. What's scarce isn't the raw data — it's the 4 percent that survives three stages of filtering. Leapmotor described this conversion rate as the "thickness" of its asset base.
"Raise the conversion rate by one percentage point, and it has the same effect as adding 10,000 km of scenario assets."
The point is that the ability to find valuable moments in data already collected matters more than simply running more test vehicles.
Scenario data comes in through four channels. In-house development and test vehicles account for the largest share, at 70 percent, followed by data collected from production vehicles already in customer use at 15 percent. Customer complaints, reviews, and interviews — voice-of-customer (VOC) data — make up 5 percent, and scenarios co-built with strategic partners, including Stellantis, account for the remaining 10 percent.
Data from these sources flows into a cloud-based scenario hub, where it is filtered and labeled under common rules before being distributed to development and validation teams. Leapmotor also runs a "T+1" system that folds newly collected data into its asset library by the following day.
VOC data, in particular, doesn't stop at customer-service records. It's sorted into six dimensions: safety, ride comfort, efficiency, energy consumption, reliability, and user experience. A complaint that sudden braking felt uncomfortable in a specific setting becomes a validation scenario — one that includes the road conditions, vehicle speed, obstacles, surrounding traffic, weather, and the system's decision at the time. Customer feedback is thereby converted into engineering parameters and test conditions.
The value of a global scenario asset comes down to how far it travels. Data collected overseas that gets used once in that one country, then discarded, only drives up the cost of local testing over time. But a situation discovered abroad that can be reproduced back in China's development environment, or in another market, turns the same data into something reused across multiple products and regions.
According to Leapmotor, 59 percent of the scenarios collected overseas were classified as transferable to other markets. This included European regulatory and traffic scenarios equivalent to 15 percent of the total, Middle Eastern heat and sand conditions at 12 percent, Southeast Asian monsoon and flooding scenarios at 18 percent, and long-distance highway conditions in the Americas at 14 percent. Roundabout and tram scenarios collected in Europe can be reproduced in China's test environment, while 50°C heat from the Middle East and flooding conditions from Southeast Asia can be incorporated into repeat testing.
The partnership with Stellantis is the channel that turns this asset movement into actual global development work. The two companies move data in both directions, jointly build scenarios, run road tests together, and carry out European regulatory and localization validation. According to Leapmotor, this two-way collaboration can expand its Europe-related scenario asset base by more than 50 percent.
"You collect it once and use it around the world."
That changes what globalization means. Instead of duplicating vehicles and development organizations country by country, the goal is to make validation knowledge gained in one region usable across markets, vehicle programs, and future projects.
Real-world roads alone can't capture every risk and every edge case. Some situations occur too rarely, or are too dangerous to reproduce on purpose. Leapmotor is building a system that refines the scenarios it has actually collected, recombines their conditions, and uses AI to generate new test scenarios from them.
Leapmotor described five generation methods: operating-condition perturbation, boundary-condition perturbation, fault injection, aging simulation, and composite-condition generation. It can vary speed, gradient, and obstacle placement, inject sensor failures and communication dropouts, or combine night driving, heavy rain, construction zones, and sensor faults into a single scenario. Leapmotor is using a limited set of real scenarios as seed data to widen the range it can actually test — going, as it put it, "from 10 percent of real scenarios to 100 percent of the edge-case range," with a goal of expanding its validation modules tenfold.




Leapmotor draws its global scenario data from in-house development and test vehicles (70%), production-vehicle samples (15%), voice-of-customer input (5%), and scenarios co-developed with strategic partners (10%). The data enters a cloud-based scenario hub and is added to the asset library on a T+1 basis.



Scenarios become evidence for market entry

Validation follows a V-model closed loop, moving from requirements and scenario construction through data validation, simulation regression, proving-ground tests, and vehicle-level testing.
At each gate in Leapmotor's process, a feature can't move to simulation unless it meets the scenario library's requirements. It can't proceed to proving-ground validation without clearing simulation regression, and only features that complete closed-loop validation on the proving ground move into production. The idea is to catch problems in data and simulation before they surface late in development or during vehicle-level testing.
This isn't only a question of development efficiency. It touches a regulatory question — what, exactly, proves that an autonomous vehicle is safe enough for the market. Shi pointed to China's GB 44721-2026 safety requirements for autonomous driving systems and the UN WP.29 work on a global technical regulation for ADS. In this emerging regulatory environment, manufacturers will increasingly need to show not only that testing was completed, but also which scenarios were validated, under what conditions, and how identified problems were resolved and traced.
Leapmotor laid out a roadmap: in 2026, build out testing standards, an integrated scenario platform, and a global collection network, and raise the share of digitalized testing to 80 percent; in 2027, apply AI-generated scenarios and a world model to validation work and reach a 60 percent asset-reuse rate; by 2028, contribute its methodology to industry standards and expand it into a global validation capability.
"What we have to do is turn data into assets, turn assets into capability, and turn that capability into a standard."
What Shi's presentation showed wasn't the scale of 1 million km of testing. The number that matters is the 40,000 km left after filtering — and how much of it can be reused across markets, vehicle programs, and future projects.
The next stage of globalization does not call simply for more distance driven. It requires a system that can retrieve and reuse a scenario wherever it is needed. But a validated vehicle does not become a viable business on its own. The next challenge is turning that technology into a service supported by day-to-day operations.



According to Leapmotor, 59 percent of the scenarios collected overseas were classified as transferable to other markets, including regulatory and traffic scenarios from Europe, heat and sand conditions from the Middle East, monsoon and flooding scenarios from Southeast Asia, and long-distance highway conditions from the Americas.   

AEM(오토모티브일렉트로닉스매거진)