From Model Development to Deployment: Ray's Computing Resource Scheduling Practice in Xiaohongshu AI Data Production | AICon Shanghai

Background

Li Auto has historically been perceived as a company focused on extended-range electric vehicles (EREVs) and cabin experience rather than cutting-edge autonomous driving. However, facing intensifying competition and evolving consumer expectations, the company is now making a major bet on intelligent driving as a core differentiator. The Turing 2.0 system marks its transition from relying on external suppliers to full self-development, a significant and resource-intensive undertaking.

Key Points

1. Technical Architecture: Full-Stack Self-Development and End-to-End Model

• End-to-End Neural Network: The system is built on a single, large end-to-end model that processes sensor inputs directly to output driving decisions. This is a departure from traditional modular pipelines and is considered a more advanced, albeit challenging, approach to autonomous driving.
• Sensor Suite: It integrates a comprehensive sensor suite, including a forward-facing 128-channel LiDAR, 11 cameras, and millimeter-wave radars. The emphasis on high-resolution LiDAR is a strategic choice for enhanced perception, especially in complex urban environments.
• Computational Hardware: The system runs on the NVIDIA DRIVE Orin-X chip, providing the necessary computational power for the end-to-end model.

2. Core Capability: Nationwide, All-Scenario NOA

• The headline feature is the promise of a full-scenario NOA that can operate on all road types in China—from expressways and ring roads to complex urban streets and narrow alleys—without relying on high-definition maps.
• This "mapless" approach is crucial for scalability, as it removes the dependency on costly and constantly updating HD map data, enabling the system to function anywhere.

3. Strategic Rollout and Application

• The rollout is planned in phases, with the system first activating on the newly launched Li Auto MEGA MPV.
• A major update is scheduled for Q4 2025, which will enable the nationwide all-scenario NOA function for eligible models.
• The system is branded as part of Li Auto's broader "Turing" platform, indicating it is designed to be scalable across future vehicle models.

Significance

A Competitive Necessity and Market Repositioning: This launch is Li Auto's direct response to the "smart driving arms race" in China's EV market. By developing a system comparable to Xpeng's XNGP and Huawei's ADS 2.0, Li Auto aims to shed its image as a "smart cockpit" leader without matching autonomous driving prowess and compete on a more holistic technology platform.

A High-Risk, High-Reward Bet: Developing a full-stack, end-to-end system from the ground up is extremely capital and talent-intensive. Success could propel Li Auto into the autonomous driving vanguard; failure or significant delays could cede ground to more established competitors and impact its premium brand positioning.

Focus on Pragmatic Consumer Value: The nationwide NOA capability directly addresses a key consumer pain point—the lack of usable advanced driver-assistance in diverse driving conditions. By targeting a "drive everywhere" use case by the end of 2025, Li Auto aims to offer a tangible, marketable advantage that could justify vehicle pricing and enhance customer loyalty.

Long-Term Foundation for Future Mobility: The Turing platform is not just for the next car model. It represents Li Auto's foundational investment in software-defined vehicle technology. Mastering end-to-end AI driving models is essential for any future ambitions in robotaxis or more advanced autonomous services, making this a critical step in the company's long-term evolution.