Emilio Frazzoli, CTO, nuTonomy - MIT Self-Driving Cars

Emilio Frazzoli's Background and Neutonomy's Origin

• Frazzoli is the CTO of Neutonomy and inventor of the RRT* algorithm; formerly a 15-year faculty member at MIT directing the research group that deployed the first autonomous vehicles in Singapore.
• Neutonomy originated from a 2009 project proposal on future urban mobility in Singapore, initiated after Frazzoli articulated a vision for autonomous vehicle sharing similar to smartphone ride-hailing apps.
• The company focuses on Level 4 and Level 5 automation, prioritizing fully driverless operation over intermediate levels of partial automation.

Economic and Societal Value of Autonomous Mobility

• Frazzoli estimates the annual societal harm of road accidents in the US at nearly $1 trillion ($300 billion economic cost + $600 billion pain/suffering), with the "cost of statistical life" valued at approximately $9 million.
• The economic value of reclaiming driver time in the US is calculated at approximately $1.2 trillion annually, exceeding the value of accident prevention.
• The potential value of a fully functional vehicle-sharing model is estimated at $2 trillion annually, based on a "sharing factor" where one shared vehicle substitutes for four privately owned vehicles.
• Key barriers to vehicle sharing adoption (current friction) include vehicle availability and parking spot availability, which autonomous vehicles can solve by operating without drivers.

Critique of Automation Levels and Industry Strategy

• Frazzoli opposes the sequential progression through SAE Levels 0–5, arguing that Levels 2 and 3 (requiring human supervision) create dangerous failure modes like mode confusion and loss of situational awareness, similar to early aviation autopilot issues.
• He contends that true value capture (safety, time reclamation, and car sharing) is only possible with Level 4 or 5 automation where no human intervention is required.
• The industry is split between two paths:
  • OEM Path: Automakers adding features to consumer cars, aiming to sell Level 2/3 autonomy packages for $10k–$20k (constrained by the net present value of the driver's time).
  • Service Path (Neutonomy/Waymo/Uber): Building dedicated fully autonomous fleets starting in geofenced areas, scaling up operations without requiring human intervention.

Technical and Operational Advantages of the Service Model

• Cost Structure: In the service model, the automation package can cost up to $100,000 because it is compared against the cost of three human drivers required for 24/7 service ($100k/year), whereas consumer models are capped by the $20k value of driver time.
• Mapping: HD maps are less of a barrier for service providers; a fleet of 1,000 cars can generate and maintain maps continuously, making map costs scale with the square root of the customer base.
• Maintenance: Professional fleets can manage sensor calibration and software updates, removing the burden from individual consumers.
• Job Market Impact: Mobility is "power limited" (lack of willing drivers); autonomous vehicles increase mobility supply rather than causing mass unemployment, though driver wages may decrease.
• Safety Statistics: Trucking is the most dangerous industry in the US, with 25% of job-related deaths; automation could improve this by allowing remote supervision.

State of Autonomous Technology and Driving Policy

• Frazzoli critiques current demos for lacking significant progress over 20-year-old technology (e.g., Ernst Diekmann's 1990s highway driving).
• Current state-of-the-art involves real-world operations in complex urban environments (e.g., Singapore, Boston Seaport) handling construction zones, jaywalkers, and unpredictable traffic interactions.
• He rejects "end-to-end" deep learning approaches for driving policy, citing risks of learning incorrect behaviors (e.g., speeding up on yellow lights) and lack of explainability/verifiability.
• Core Philosophy: Road rules are designed to minimize negotiation; therefore, autonomous systems should rely on formal verification of rules rather than rule generation via data.
• Solution Approach: Neutonomy uses sampling-based planning (RRT*) to generate trajectories and checks them against a hierarchy of formal rules to find the "minimum violation" solution.
• Hierarchy of Rules: Critical rules (do not hit humans) are weighted higher than secondary rules (stay in lane), allowing the system to violate minor rules to satisfy major constraints.

Unresolved Challenges in Rules and Ethics

• The primary challenge identified is the lack of a rigorous, mathematical theory for current human driving rules, which are often vague (e.g., "do not pose an obstacle") or contradictory.
• Frazzoli highlights the difficulty of "right-of-way," noting there is no single mathematical definition for it in current traffic laws.
• Ethical dilemmas (Trolley Problems) in autonomous driving involve probabilistic trade-offs (e.g., swerving to avoid a jaywalker with probability $p$ of hitting a bystander), which require community-defined policies rather than purely algorithmic solutions.
• The industry must first develop a sound theory for human driving behavior to effectively program autonomous vehicles, as compliance with precise rules will drive sensing and perception requirements.

Future Outlook and Company Growth

• Neutonomy expects rapid adoption of autonomous "mobility-as-a-service" fleets within a couple of years, while consumer ownership of fully autonomous cars is projected 20+ years away.
• The company plans to double its size and hire several hundred people in the next two years to support this expansion.