Vision–Language–Action
Unified perception, reasoning, planning, and control for complex real-world environments.
Foundation models for embodied intelligence
Intelligence that can perceive, reason & act.
I am Kun Zhan, Head of Foundation Models and Autonomous Driving at Li Auto. I build vehicle-scale AI systems that connect perception, language, decision-making, and action—and carry them from frontier research into production.
Beijing / San Jose Li Auto Autonomous Driving · Foundation Models · Embodied AI
詹锟 Building physical-world intelligence from research to road
Research snapshot · Jul 2026
Production research across autonomous driving, multimodal AI, and
world models.
Publications
63
Citations
2,068
h-index
18
i10-index
29
01 / About
From machine intelligence to language intelligence—and back to action.
My work focuses on unifying the capabilities a physical agent needs: understanding three-dimensional scenes, reasoning about intent and risk, planning under uncertainty, and executing safely in real time.
Mission
Build physical-world AGI, starting with autonomous driving and expanding toward robots and intelligent spaces.
World models & reinforcement learning
Simulation, generative scene models, closed-loop evaluation, and learning from physical feedback.
Model–system co-design
On-vehicle inference, model–chip co-design, data engines, and reliable fleet-scale deployment.
02 / Milestones
A path from prediction systems to vehicle-scale foundation models.
Selected moments across leadership, production systems, model releases, and research.
Foundation-model leadership
Leading Li Auto's unified foundation-model and autonomous-driving agenda across Mach VLA, Mach Mind, world models, reinforcement learning, infrastructure, and deployment.
Livis Day keynote
Presented Li Auto's software and embodied-intelligence roadmap, connecting intelligent vehicles with a broader physical-world AI stack.
World models & CVPR research
Contributed to ReconDreamer, StreetCrafter, and DrivingSphere—advancing reconstruction, controllable scene generation, and closed-loop 4D simulation for autonomous driving.
DriveVLM & ECCV research
Released DriveVLM and contributed to Street Gaussians and TOD3Cap—bridging multimodal reasoning, dynamic urban reconstruction, and 3D scene understanding.
Joined Li Auto
Helped evolve the driving stack from Highway NoA and City NoA through end-to-end, VLM-assisted, and VLA-based architectures running across production vehicles.
Baidu Apollo
Led L4 prediction and pre-decision algorithms for robo-taxi pilots and production-oriented autonomous-driving systems.
03 / Experience
Research leadership grounded in shipping real systems.
Li Auto
Apr 2021 — Present · Beijing / San JoseHead of Foundation Models & Autonomous Driving
- Lead foundation-model and autonomous-driving teams spanning VLA, agentic LLM/VLM systems, world models, reinforcement learning, data infrastructure, and on-vehicle deployment.
- Drive model–chip–OS–controller co-design and production integration across Li Auto's full-stack AI platform.
- Guide a 100+ person organization across perception, planning, foundation models, simulation, data, and deployment.
Site Manager, U.S. R&D Center
- Launched Li Auto's overseas research hub and connected Silicon Valley research with Beijing execution.
Baidu Apollo
Apr 2016 — Mar 2021 · BeijingAlgorithm Lead, L4 Prediction & Planning
- Led prediction and pre-decision algorithms for L4 robo-taxi pilots in complex urban traffic.
- Delivered planning-and-control modules and onboard deep-learning components for fleets in Beijing and Guangzhou.
Education
2009 — 2016Beihang University · M.S. in Navigation, Guidance and Control
Research focused on object recognition and tracking.
University of Science and Technology Beijing · B.Eng. in Automation
04 / Selected research
Research that expands how vehicles see, imagine, and decide.
Selected work across VLM/VLA systems, world models, 3D reconstruction, planning, and simulation.
VLM · 2024
DriveVLM: The Convergence of Autonomous Driving and Large Vision-Language Models
Combining multimodal reasoning with driving perception, planning, and interaction.
3DGS · ECCV 2024
Street Gaussians
Modeling dynamic urban scenes with Gaussian splatting.
World model · CVPR 2025
ReconDreamer
Crafting world models for driving-scene reconstruction via online restoration.
Video diffusion · CVPR 2025
StreetCrafter
Street-view synthesis with controllable video diffusion models.
4D simulation · CVPR 2025
DrivingSphere
Building a high-fidelity 4D world for closed-loop simulation.
05 / Updates
A living window into current work.
Short notes for model releases, talks, research, and milestones—without the overhead of a full blog.
Let’s build what moves intelligence forward.
For research conversations, speaking, advisory work, or collaboration, reach out directly.