Li Auto i8’s AI Chauffeur: VLA Tech Will Steers Family EVs into Autonomy Era
How China’s NEV leader is betting on embodied AI to win the autonomy race
As a large-sized pure electric SUV, the Li i8 adopts a three-row, six-seat layout. Li Auto has achieved best-in-class levels for both the second-row center aisle width and third-row legroom. However, this passenger comfort comes at the expense of cargo space. With all six seats occupied, the trunk appears insufficient for six passengers’ luggage.
Compared to other pure EV brands, the i8 takes a different approach to configurations. It offers three variants, all equipped as standard with the 5C ultra-fast charging battery and dual-chamber air suspension. This is a significant attraction for consumers prioritizing comfort but concerned about range anxiety.
Yet, space and configuration aren’t the i8’s main highlights. Li Auto seems to be betting heavily on its VLA (Vision-Language-Action) large model-powered advanced driver-assistance system (ADAS).
Will the VLA Large Model Be Li Auto’s New Growth Space?
AI large models featured prominently at the Li i8 launch event, signaling the sales-leading new energy vehicle (NEV) brand’s clear intent to transform into an AI company. Within this strategy, the VLA large model for ADAS represents a crucial bet on the future.
VLA is a new technical approach for ADAS proposed by Li Auto. Unlike the mainstream end-to-end method, VLA enables self-upgrading capabilities. It leverages AI large models to expand data volume and diversity, allowing the ADAS to learn from more obstacle scenarios and traffic rules.
Li Auto believes VLA can perceive the physical world comprehensively through a combination of 3D and 2D vision, unlike Vision-Language Models (VLM) which primarily parse 2D images. Simultaneously, VLA possesses a complete “brain system” with language and Chain-of-Thought (CoT) reasoning capabilities. It can see, understand, and execute actions, aligning with human cognitive processes.
Technically, VLA holds promise for addressing limitations of the end-to-end approach: limited data collection often hinders handling edge cases effectively.
However, different technical paths face distinct challenges. VLA represents uncharted territory in ADAS technology, offering Li Auto no precedent to follow. From self-developed models to data training, Li must forge its own path. Consequently, the industry watches closely how Li develops VLA and its real-world performance.
A week before the i8 launch, Li Auto held a media briefing in Beijing. Lang Xianpeng, Senior Vice President of Autonomous Driving R&D, and core team members engaged with media, providing comprehensive answers on VLA’s development path and performance.
Q&A Highlights (Adapted):
Q: VLA demonstrates reasoning capabilities and more human-like behavior, but requires seconds for reasoning. How does it handle “fast thinking” in sudden scenarios?
Lang Xianpeng: The perceived slowness is primarily due to visualization. The actual reasoning speed is fast; we display key steps for clarity. VLA’s reasoning frame rate is around 10Hz, over three times faster than our previous VLM (Vision-Language Model) at ~3Hz.
Q: You mentioned using MindGPT as the base model now. Is it better?
Zhan Kun: Our self-developed base architecture is crucial for deploying VLA. VLA is a 4B model, larger but faster. This efficiency stems from our proprietary architecture – a customized MoE (Mixture of Experts) 0.4×8 design, unique in the industry and optimized for embedded chips.
Q: How do you judge the timeline for autonomous driving (AD) rollout and commercialization?
Lang: Technically, VLA can progress towards higher-level autonomy, but it’s nascent. Currently, VLM performance ≈ the upper limit of end-to-end. Progress won’t be slow; end-to-end MPI jumped from 10 to 100 in ~1 year. VLA could reach 1000 MPI soon. Core enablers are algorithm, computing power, data, and engineering. Commercialization heavily depends on regulations, insurance, liability frameworks.
Q: What are VLA’s core challenges and prerequisites for companies?
Lang: You cannot skip foundational steps like rule-based algorithms or end-to-end development. VLA requires a robust data foundation (we have 1.2 billion km), substantial computing power (training & inference), and engineering capability. Without accumulated data and infrastructure, training a world model is impossible.
Q: Competitors offer 2000 TOPS compute. What are Li Auto’s plans?
Lang: Effective compute depends on quantization precision. We have long-term plans but can’t disclose specifics now.
Q: How does VLA prioritize between the “impossible triangle” (Efficiency, Comfort, Safety)?
Lang: Safety (MPA – Miles Per Accident) is paramount. Our target is 10x safer than human drivers (6 million km/accident). Comfort is next; poor comfort causes disengagements. Efficiency comes last. We prioritize safety and comfort over risky maneuvers for efficiency.
Q: Why drastically reduce real-world testing (RWT) to 20,000 km?
Lang: RWT is costly, hard to reproduce edge cases, and inefficient. Our simulation now matches RWT reliability (>90% of VLA testing is simulation). Simulation is faster, cheaper, and crucial for VLA’s reinforcement learning. Hardware testing remains essential.
Q: Was VLA the only considered path beyond end-to-end?
Lang: VLA, representing Embodied AI, was the most promising path identified through extensive research. It’s a framework applicable beyond AD to future robotics.
Q: How does Li Auto maximize chip utilization (e.g., Thor FP8)? Will you develop your own chips?
Zhan Kun: Our engineering team has deep expertise in optimizing deployment (e.g., on Orin, now Thor). We modified CUDA, rewrote PTX instructions. VLA inference sped up 10x (from 500ms/frame to 10Hz). Techniques like FP8/FP4 quantization are key. Custom chips become viable when algorithms stabilize for cost/efficiency.
Q: Is VLA mainly an engineering innovation?
Zhan Kun: No. VLA is an algorithmic innovation applying Embodied AI principles to AD. It’s a true end-to-end model integrating perception, reasoning (Language), and action. The challenge is deploying large models on edge devices, requiring significant engineering optimization.
Q: How do you balance model size and inference efficiency for VLA deployment?
Zhan Kun: We use a unique 8×0.4B MoE architecture suited to Nvidia chips. We distilled a 32B cloud model into a 3.2B MoE model. Optimizations like Flow Matching reduced diffusion steps from 10 to 2, enabling deployment.
Q: How did VLA learn precise commands like “move forward 10 meters”? Can we trust its judgments?
Zhan Kun: The model learns spatial understanding (distances) from vast data. It exhibits compositional generalization – combining concepts (numbers, units, actions). Trust is built via rigorous data cleaning, targeted data generation, and techniques like Super Alignment for compliance. RAG (Retrieval-Augmented Generation) enhances memory/planning.
Q: How do you achieve 99.9% simulation accuracy?
Zhan Yi Fei: We validated our world model simulation against >1.5 million km of 2024 RWT data. We spent a year fixing discrepancies (e.g., traffic light visibility at 200m) to achieve high fidelity.
Q: How does OTA 7.5’s Super Alignment help VLA?
Lang: Super Alignment improved data quality/coverage and added compliance rules. Crucially, it built a massive library of test scenarios (>400,000) essential for rapidly evaluating VLA iterations. Daily simulation peaks reached 300,000 km.
Q: Will VLA enable personalized driving styles?
Lang: Yes, reinforcement learning enables adapting to user style over time, a capability end-to-end lacked. On-device reinforcement training is being considered.
Q: How do you maintain accuracy reducing precision (FP16 -> FP8 -> FP4)?
Zhan Kun: Large models are more tolerant of lower precision (FP8/INT8). Key is Quantization-Aware Training (QAT) – training the model *with* target precision constraints. Extensive data cleaning is vital for stability. FP4 requires even cleaner data and stable training. On Orin (no FP8), we use INT4 + mixed precision. FP4 on Blackwell could enable 20Hz.
Q: What perception improvements come with VLA?
Lang: VLA brings significant upgrades: dynamic object detection range extended to 200m (vision-only), OCC (Occupancy Network) range to 125m. We continuously enhance core capabilities.
Q: Biggest challenges/mistakes in developing VLA? Organizational changes?
Lang: The biggest challenge was transforming the R&D process for reinforcement learning – building reliable world models and efficient simulation. Team size grew modestly (End-to-End: 180, VLA: 200+). We focused on algorithm, data, and computing power. Avoiding major strategic errors was key. Data foundation was paramount.
Q: How do you prevent model hallucinations/anti-commonsense outputs?
Zhan Kun: Rigorous data cleaning, targeted data generation to fill knowledge gaps, and Super Alignment techniques ensure outputs are commonsensical and compliant.
Q: How do you define “high-quality data” with simulation?
Zhan Yi Fei: We seek “expert driver data.” Cloud-based programs and models screen data against “expert” criteria (safety, comfort, compliance, efficiency). Context matters (e.g., navigating around illegally parked cars).
Lang: Post world-model training, generated data supplements corner cases. For reinforcement learning, data primarily trains the world model itself. Generated data is increasingly crucial for breadth and depth.
Q: Closing thoughts on VLA’s future?
Lang: We started AD data collection early (since Li ONE). Our 5-year data/engineering foundation is a significant moat. While VLA seems incremental now, its scaling potential is immense. A mature world model enables “training in simulation” at unparalleled speed (like an agent in SimCity). Just as end-to-end accelerated progress, VLA will drive a leap within a year. Li Auto is positioned to lead this transformation.