Xia Zhongpu, the executive who led development of Li Auto’s end-to-end (E2E) assisted driving model, has left the company, signaling a broader strategic shift at the Chinese automaker.
Xia, who joined Li Auto in 2023 and reported directly to Lang Xianpeng, vice president of assisted driving R&D, was a prominent figure within the company’s advanced driving unit. People familiar with the matter said he had already stepped away from Li Auto’s vision-language-action (VLA) model development team and had not attended internal meetings for several weeks. His next move remains unclear.
Li Auto did not respond to 36Kr’s request for comment.
Xia’s departure marks a turning point in the company’s evolving approach to assisted driving technology. He was instrumental in Li Auto’s early E2E architecture, which gained traction following promising results and led to a restructuring of the assisted driving division in November 2024. The company established three dedicated teams at the time: E2E modeling, world modeling, and production R&D. Xia was named head of the E2E unit.
During his tenure, Xia rose quickly within Li Auto’s hierarchy. However, internal sources suggest his exit may stem from differing views on the company’s technological direction. “Xia believed the E2E approach could still be optimized,” said one person familiar with the matter. “But internally, the company had already gone all-in on VLA.”
The transition was underscored on May 7, when CEO Li Xiang described VLA during a company-hosted “AI Talk” session as “a large model that drives like a human.” He also revealed that Li Auto had tripled its planned GPU training allocation to support VLA development.
This pivot has coincided with a reshuffling of leadership. Lang has reportedly been promoted, and Jia Peng, previously in charge of the company’s world model framework, now leads the VLA project as head of autonomous driving technology R&D.
Li Auto’s assisted driving roadmap has undergone multiple changes since 2023. The company moved away from high-definition maps and rule-based logic, adopted E2E frameworks, and now appears committed to VLA as its core development path.
Originally popularized by Tesla, the E2E model allowed Li Auto to streamline core modules—perception, prediction, planning, and control—into a unified system driven by machine learning. The company formally began E2E development in November 2023 and rolled out a combined E2E and vision-language model (VLM) to Max users by October 2024, reportedly two months ahead of schedule.
That rollout was seen as a major win, helping Li Auto shed its image as a laggard in assisted driving and emerge as a top-tier contender. Xia, as the head of the production-ready E2E system, benefited from the momentum and earned internal promotions.
Yet, consensus on E2E’s long-term viability never fully materialized. At the same May 7 event, CEO Li Xiang noted a critical limitation: “End-to-end doesn’t fully understand the physical world. It’s more of an imitation process,” he said, pointing out that the system performed well in general scenarios but struggled in complex, unfamiliar ones.
Management saw those constraints as fundamental. While Li Auto had already added VLM features to its E2E stack, the team concluded that vision-language alone could not close the gap in higher-order decision-making.
VLA, originally developed by Google DeepMind for robotics applications, is now viewed within Li Auto as the future of embodied intelligence. Unlike VLM systems that parse 2D visual inputs, VLA incorporates an action component, allowing models to not only interpret their environment but also perform physical tasks like driving. These systems support 2D and 3D vision, language comprehension, and chain-of-thought reasoning, essentially enabling vehicles to see, think, and act more like humans.
Industry peers are also trending in this direction. Nio has released a world model focused on reading road signs and textual input, while Xpeng has introduced a foundation model with advanced reasoning capabilities that has since been optimized for in-car use.
Still, the VLA route is not without skeptics. Some experts interviewed by 36Kr noted that the approach remains largely untested in real-world conditions. Even Li acknowledged the risks: “In truth, we’re venturing into uncharted territory,” he said.
More than just a personnel change, Xia’s exit may perhaps reflect the end of an ambitious phase in Li Auto’s assisted driving journey, and the company’s deeper bet on a new frontier.
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This article was written by Li Anqi for 36Kr.