Automotive
How Li Auto Doubled ADAS Simulation Coverage Using NVIDIA Omniverse NuRec
Objective
Li Auto is one of China's leading intelligent electric vehicle manufacturers with a full-stack, self-developed Advanced Driver Assistance System (ADAS) platform. At the center of its development process is an internally built, closed-loop simulation environment for algorithm evaluation, training, and regression validation against real-world driving data. The platform reconstructs real driving logs into 3D simulations, then replays those scenes repeatedly to evaluate how its ADAS algorithms respond before any update reaches the fleet. Validation coverage scales directly with how many scenes it can reconstruct. That makes reconstruction cost and throughput a direct ceiling on how much of its fleet data becomes usable for algorithm testing.
To scale closed-loop ADAS simulation to production while controlling per-scene reconstruction costs, Li Auto is integrating NVIDIA Omniverse™ NuRec into this platform. NuRec converts real driving logs into high-fidelity 3D scenes, enabling reproducible closed-loop validation without proportional increases in compute.
Customer
Li Auto
Use Case
Simulation / Modeling / Design
Topic
Simulation / Modeling / Design
Products
Key Takeaways
Simulation Scale
- 2x simulation scene coverage on equivalent compute, extending algorithm validation across a broader range of driving and parking scenarios.
Reconstruction Efficiency
- ~50% reduction in per-scene reconstruction cost through 3DGUT's unified camera and LiDAR representation; 4x multi-GPU training speedup on 8 GPUs, improving throughput for incident reconstruction and simulation iterations.
Simulation Quality
- Improved asset completeness and visual consistency in sparse-viewpoint scenarios through TokenGS and Harmonizer, yielding more reliable simulation inputs for ADAS model training.
What Makes Multimodal ADAS Closed-Loop Simulation Hard to Scale?
Scaling closed-loop simulation to production quality across multimodal sensors required addressing three compounding constraints:
Multimodal Reconstruction and Rendering
LiDAR point clouds are a primary input for ADAS algorithms alongside camera feeds. Conventional 3D Gaussian Splatting approaches are not optimized for LiDAR data. Maintaining separate camera and LiDAR reconstruction models increases engineering overhead and makes multi-sensor simulation consistency harder to achieve.
Reconstruction Throughput
Processing driving logs at production scale demands high reconstruction throughput. Long per-scene reconstruction times create bottlenecks in incident replay, algorithm regression, and closed-loop simulation iteration speed.
Novel View Synthesis Quality
Driving log scenes cover large spatial extents with relatively sparse viewpoint coverage. Without training signals for unseen camera angles, reconstruction models tend to overfit to captured viewpoints rather than forming stable, generalizable 3D representations needed for reliable simulation.
How NVIDIA Omniverse NuRec Enables Production-Scale Neural Reconstruction
Li Auto developed its internal reconstruction solution by integrating capabilities from NVIDIA Omniverse NuRec into its proprietary closed-loop simulation platform. NuRec converts raw driving log data into reconstructed 3D scenes and generates high-fidelity sensor renders for both camera and LiDAR. Four NuRec capabilities were integrated into the pipeline:
3DGUT: Unified LiDAR and Camera Representation
NuRec's 3D Gaussian with Unscented Transforms (3DGUT) representation allows a single set of Gaussians to model and render both LiDAR point clouds and image frames simultaneously. This eliminates the need for separate camera and LiDAR pipelines, reducing engineering overhead and improving cross-modal simulation consistency.
Hybrid Multi-GPU Training
A combined DDP and FSDP parallel training strategy distributes 3D Gaussian particles across multiple GPUs. This approach accelerates reconstruction of large-scale ADAS driving scenes, with targeted support for dynamic and deformable objects. At an 8-GPU configuration, the approach delivers approximately 4x acceleration over a single-GPU baseline.
TokenGS Asset Generation (Asset Harvester Workflow)
The Asset Harvester pipeline converts sparse object observations from real driving logs into more complete 3D assets, improving visual quality in novel viewpoints. Li Auto trained and deployed its own TokenGS model to support this asset generation workflow on its internal dataset.
Harmonizer Visual Harmonization
NuRec's diffusion-based Harmonizer model corrects artifacts in novel-view synthesis and improves consistency in appearance, color, and shadow between inserted assets and the NuRec-reconstructed scene background. The result is visually coherent closed-loop simulation scenes suitable for ADAS algorithm training and evaluation.
How NVIDIA Omniverse NuRec Establishes a Durable ADAS Simulation Foundation
Li Auto’s implementation demonstrates that NVIDIA Omniverse NuRec serves a function beyond single-scene reconstruction efficiency—establishing a durable infrastructure investment. As NuRec capabilities continue integrating into Li Auto’s closed-loop workflow, the platform is positioned to increase ADAS development velocity, improve simulation fidelity, and expand test coverage without requiring equivalent growth in physical compute resources.
As an active participant in production-scale ADAS development, Li Auto is using simulation infrastructure to advance algorithm validation and training across its deployed fleet. Continued integration of NuRec capabilities into the closed-loop pipeline provides a foundation for improving both the pace and quality of intelligent driving system iteration.
NVIDIA Omniverse NuRec is built on a modular software architecture that supports different integration patterns. It can function as a complete end-to-end neural reconstruction solution, or developers can select individual components, including the reconstruction framework, rendering engine, and generative model modules, for integration into custom workflows. Open source distribution of NuRec allows teams to adopt, extend, and contribute back to the platform as requirements evolve.
"Closed-loop world model simulation is a foundational capability in Li Auto's intelligent driving R&D. We build and own our simulation platform to match our own technical architecture. In that process, we worked closely with NVIDIA on neural reconstruction, multi-GPU parallel training, asset generation, and visual consistency. Integrating relevant NuRec capabilities into key stages of our pipeline improved scene reconstruction efficiency and multimodal simulation quality, accelerated validation and production deployment of related technologies, and contributed to our ability to continuously evolve our autonomous, scalable closed-loop simulation infrastructure."
Zhan Kun
Head of Foundation Models, Li Auto
Learn more about how NVIDIA powers simulation and validation for autonomous vehicles.