On the Road With DRIVE AV

· Source: NVIDIA · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Software Development & Engineering · Depth: Advanced, long

Summary

The Alpamayo autonomous vehicle (AV) stack, developed by NVIDIA, features a unique hybrid architecture combining an end-to-end deep learning model with a traceable, human-engineered classical AV stack. This dual approach aims to achieve human-like driving capabilities while ensuring safety through the classical stack's guardrails, which prevent the end-to-end model from operating outside its distribution. The system demonstrates rapid development iteration, with approximately 2,300 versions over a year, averaging seven model iterations daily. This speed is attributed to the end-to-end model's ability to backpropagate improvements without introducing regressions seen in modular classical systems. The hybrid stack is designed for scaling to L3/L4 autonomous driving and robotaxi services, supported by teleoperations for backup and advanced AI tools like NuRec and Cosmos for data curation and synthetic scenario generation. The system has achieved the highest NCAP safety rating, showcasing its robust safety and performance in complex urban and highway environments, including challenging lane changes and obstacle avoidance.

Key takeaway

For computer vision engineers developing autonomous driving systems, adopting a hybrid AV stack like Alpamayo's offers a path to combine the agility and human-like performance of end-to-end models with the verifiable safety of classical systems. You should prioritize integrating a robust classical guardrail to manage the inherent verification challenges and potential regressions of deep learning models, enabling faster iteration and confident deployment of L3/L4 capabilities.

Key insights

A hybrid AV stack combining end-to-end AI with a classical safety system offers both human-like driving and robust safety.

Principles

Method

The Alpamayo AV stack integrates an end-to-end deep learning model for human-like driving with a classical, ASIL-protocol-compliant stack that acts as a safety guardrail, preventing out-of-distribution behavior and ensuring safety during rapid model iteration.

In practice

Topics

Best for: Computer Vision Engineer, AI Engineer, Machine Learning Engineer, Robotics Engineer

Related on AIssential

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by NVIDIA.