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

The primary reasons cited are increased safety (reducing accidents caused by human error), convenience (allowing passengers to do other activities), improved mobility access for various groups, increased traffic efficiency, and reduced environmental impact.

While autonomous vehicles are expected to be safer asymptotically, demonstrating this statistically is challenging. The return of time value to society from not driving ($1.2 trillion) significantly outweighs the economic cost of road accidents ($1 trillion). Furthermore, efficient car sharing enabled by autonomy could provide an additional $2,000 per year per user.

Levels 2 and 3 automation require human supervision and the ability to intervene, which can be dangerous due to human nature and lack of training. Levels 4 and 5 offer full or high automation where the car drives itself without needing human intervention, which Frazzoli argues is essential for truly capturing the value of the technology.

For services like ride-sharing to be truly convenient and reliable, cars need to operate without a human driver, allowing them to pick up and drop off passengers autonomously. This capability is not possible with lower levels of automation, which are seen more as 'nice gadgets'.

Selling as a product limits the autonomy package cost to a few thousand dollars and requires operation everywhere. As a service, the cost constraints are looser because it competes with human drivers, and the operational scope can be limited to specific geofenced areas and conditions, making the problem much more manageable.

Frazzoli suggests the impact is more on increasing mobility supply rather than widespread job loss. He points out that many mobility services are currently manpower-limited. While driver wages might decrease, this could be balanced by increased service availability and potentially safer conditions, especially for truck drivers.

The primary challenges are sensing and perception, HD mapping, and 'driving policy' or decision-making. While sensing and mapping are progressing, the complex decision-making in unpredictable urban environments remains a significant hurdle.

Frazzoli argues against purely learning-based systems due to the risk of learning incorrect behaviors and the difficulty in explaining or debugging decisions. He advocates for rule-based systems, enhanced with formal methods and algorithms like RT-Star, which allow for verification and more predictable outcomes.

The RT-Star algorithm generates a vast graph of potential trajectories and then checks them against a set of formal rules. This approach differs from pure generative models by efficiently verifying candidate solutions, making it easier to debug and ensure compliance with road regulations.

The biggest challenge is not just the technology, but our lack of a precise, rigorous understanding of how we want vehicles—both human-driven and autonomous—to behave. Developing a sound theory for the rules of the road and defining these behaviors precisely is crucial.