Why Xpeng Is Taking Tesla's Controversial Approach To Autonomy

"At the IAA Mobility 2025 show in Munich, Candice Yuan, senior director and head of product at Xpeng's Autonomous Driving Center, told CarNewsChina that the company has grown increasingly confident in its vision-based approach since pulling lidar from its vehicles. "The lidar data can't contribute to the AI system," Yuan told the outlet, adding that the company's large language model is fed mainly 10 to 30-second short videos, taken from its customer vehicles, and then used to train the AI system."

""We call it VLA. Vision, language, action. Lidar data is different and can't be absorbed by the AI system," Yuan added. Xpeng's self-driving system is called Navigation Guided Pilot (XNGP). It sounds far less polarizing than Tesla's Full Self-Driving (FSD), which still requires constant driver supervision and readiness to take over at all times. Like Tesla, Xpeng is betting big on end-to-end machine learning models that it says could operate anywhere in China, at least theoretically."

Xpeng removed lidar from its vehicles and shifted to a camera-first, vision-based autonomy strategy. The company's large language model is trained mainly on 10- to 30-second short videos captured from customer vehicles. The approach is termed VLA: Vision, Language, Action. XNGP (Navigation Guided Pilot) is Xpeng's self-driving system and relies on end-to-end machine learning models aimed at broad operational coverage in China. Competitors Waymo and Zoox continue to use lidar to improve perception in low light, bad weather, and complex edge cases. Lidar training introduces higher labeling, calibration, integration costs and can require reengineering if added after initial system design.

Read at InsideEVs