AI BriefingAV 3.0 Data Pipeline: The 8-Stage Autonomous Driving Development Structure Presented by AWS and NVIDIA
·2026.11.12 12:45
Key point
As the autonomous driving competition shifts from model performance to the speed of data relearning, AWS and NVIDIA have presented an 8-stage pipeline to address this.
Details
- The core of the autonomous driving competition is shifting from model performance to the speed of relearning and validation based on proprietary data
- The AV 3.0 Physical AI data pipeline reference architecture, released by AWS and NVIDIA in March 2026, provides a blueprint for data loop configuration
- The pipeline consists of 8 stages from collection to validation, forming a data flywheel that incorporates failure cases into the next learning cycle through a retrieval-driven approach
- NVIDIA provides a cloud-to-car structure connecting data center training (DGX), simulation (RTX/Omniverse), and vehicle inference (DRIVE AGX), along with the Alpamayo and Cosmos open models
- AWS addresses infrastructure bottlenecks from data collection to model training and validation by utilizing S3, SageMaker HyperPod, EC2 GPU, and others
- Hyundai Motor Group announced a two-track strategy of pursuing mass production based on NVIDIA and developing its own model (Atria AI), presenting a data flywheel roadmap
- Part 2 covers a Blueprint Lab exercise using nuScenes data and methods for scaling to customer environments
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
NVIDIA Updates Cosmos Framework and NIM for Physical AI with 2-3x Inference Speedups
AWS Releases Guide for Building Physical AI Training Pipelines Using SageMaker and IoT Greengrass