# Driving Towards the Future: Deployments, Designs, and Challenges of Autonomous Vehicles Source: https://www.youtube.com/watch?v=bsMlR70D-I4 Recap page: https://rapidrecap.app/video/bsMlR70D-I4 Generated: 2025-11-25T15:03:25.577+00:00 --- ## Quick Overview The deployment of autonomous vehicles is currently in early stages, with about 2,000 truly driverless vehicles operating in North America, relying heavily on either highly engineered classical AV stacks or newer, less explainable end-to-end neural network architectures, all facing the persistent and critical challenge of reliably handling infinite edge cases. **Key Points:** - Currently, North America has roughly 2,000 truly driverless vehicles deployed, while worldwide there are less than 10,000 operating. - The classical AV stack, used by Waymo and Zoox, is modular, allowing for causal analysis, but it is extremely engineering intensive, requiring large teams, like Waymo's 1500 software engineers. - End-to-end neural architectures, exemplified by Tesla's FSD (version 12+), allow for fast, data-driven optimization but suffer from a lack of explanability and regressions when updated. - A cutting-edge third design incorporates Large Language Models (LLMs) into end-to-end systems, enabling text input for instructions and queries, offering potential improvements in context understanding and 'open vocabulary' perception. - A major remaining challenge for all AI systems is achieving the $10^{-7}$ failure rate required for human-level safety, as engineers typically aim for only 99% accuracy, which is insufficient for safety-critical tasks. - The research lab discussed is developing a road hazard ontology grounded in real-world videos to systematically catalog and benchmark edge cases, finding that even models like GPT-5 recognize only about one-third of hazards (low recall). - Human drivers require 24 lifetimes of driving to experience a fatality, indicating that short-term success with an automation system is insufficient to judge its ultimate safety level. **Context:** This webinar, hosted by the Waterloo Center for Automotive Research (WAKCAR) and moderated by Michelle Van Djk, features Dr. Kristoff Teski, a Professor of Electrical and Computer Engineering at the University of Waterloo, who leads research on ensuring the safety of AI systems and driving behavior. Dr. Teski introduces the current state of autonomous vehicle deployments, contrasting the foundational software architectures being used, and detailing the significant challenges, particularly concerning safety validation and edge-case handling. ## Detailed Analysis Dr. Teski outlines three primary architectural approaches in autonomous driving: the classical AV stack, the end-to-end neural network approach, and the emerging Vision-Language-Action Model (VLA) architecture using LLMs. The classical stack, utilized by Waymo and Zoox, is modular with explicit interfaces between perception (heavy on neural networks), prediction, and manually programmed planning; its benefit is debuggability, but its drawback is being extremely labor-intensive. In contrast, end-to-end systems like Tesla FSD rely entirely on neural networks, allowing for easier optimization but lacking explanability, leading to regressions and issues with edge cases. The VLA model integrates LLMs to process sensor inputs and text, potentially improving context and allowing for instructions, though it inherits LLM risks like hallucinations. Currently, deployments are nascent, with fewer than 10,000 driverless vehicles globally, mostly operating in favorable weather conditions. Dr. Teski emphasizes that edge cases—situations not well-represented in training data, such as construction zones, flash floods, or sudden vehicle instability—remain the biggest challenge, leading to 'robotic driving' or critical failures, like the Cruise accident where the vehicle lacked accident context. He stresses that achieving human safety levels requires a failure rate of $10^{-7}$, far beyond typical 99% AI accuracy, and cautions that individual user experience is not a metric for systemic safety, noting that even current high-exposure systems still exhibit dangerous behaviors like phantom braking or running red lights. ### Current Deployment Status - North America has about 2,000 truly driverless vehicles (L4) deployed, primarily Waymo and Zoox, operating under specific, often favorable, conditions - Worldwide deployment is less than 10,000 vehicles, indicating early stages. ### Architectural Designs Compared - Classical Stack (Waymo/Zoox) relies on modularity and manual rules for planning but is extremely engineering-intensive - End-to-End (Tesla FSD) is data-driven and faster to develop but lacks explanability and suffers from regressions. ### Emerging VLA Architecture - Vision-Language-Action Models integrate LLMs to process visual data alongside text inputs for instructions and queries - This design promises 'open vocabulary' understanding but still faces foundational neural network reliability issues. ### Safety Benchmarking and Edge Cases - Human drivers have a fatality rate of about five per billion kilometers driven, requiring 24 lifetimes to see one fatality; AI must match this low rate ($10^{-7}$ failure rate) to be considered safe. - Edge cases are the biggest challenge, demonstrated by system failures in construction zones, hydroplaning, or misidentifying hazards like cows as debris. ### Research Focus on Hazard Identification - The Waterloo lab is building a road hazard ontology, sourcing real-world video data to create a taxonomy of edge cases - Testing cutting-edge AI (like GPT-5) against this taxonomy shows low recall (only one-third recognized) for hazardous conditions.