Imagine getting into a driverless Taxi sometime in the future. There would be no wheel or gas pedal, only passenger seats, because the car drives itself. Once an idea straight out of a page from a science-fiction novel, self-driving vehicles are becoming a reality. 

Many people are hesitant about the idea of being a passenger in a fully automatic self-driving car, with 73% of Americans saying they would be afraid.¹ The truth is, we’re not very good at driving ourselves, but we’ve become accustomed to it and accept the risk. Globally, there are a million fatalities and fifty million injuries each year related to automobile collisions.² 

Experts estimate that driverless cars would reduce the number of car accidents by 90 percent.³ Humans are flawed; we get tired, our attention drifts, and we can be aggressive. A driverless car would respond faster, never tire, and would be less forceful.⁴ It will take some getting used to, but after a few trips, we will let our guard down, and eventually, driverless cars may be all we know. Some experts even say that one day the conversation may shift to whether humans should be able to drive at all.⁵ 

Inspired by human nature, AI takes a learning network, gives it a goal, and a way to measure success against it. From there, they feed it with information, and it connects a model of understanding through experience.⁶ Self-driving cars are more similar to humans than we could ever imagine. 

A learning network is placed in a vehicle fitted with lasers, radar, cameras, and actuators, acting as the sensors that bring information from the environment into the network.⁷  Designers give it a goal to navigate the environment safely, a way to measure its success, and a few constraints to point it in the right direction. Then they turn the car on and train it through experience, with each move tuning the network until it can drive itself.  

Similar to the breakout playing AI discussed in Chapter 6, the vehicle does not know much when first turned on.⁸ The programmer hardcodes a few rules to point it in the right direction, but the rest of the connections result from reaching toward its goal through trial and error.⁹ As it bumps into the environment, moves that produce a high or low score carve out the patterns that will eventually enable the car to drive independently. After driving millions of road miles and simulating billions more, the self-driving network makes enough mistakes to teach itself how to drive.¹⁰ 

Humans can automate the skill to drive after about 50 hours of experience.¹¹ Driverless cars learn with a network like we do, but it’s primitive and takes much longer.¹² Self-driving networks are slow learners, taking a crude approach to tuning a network by feeding it with endless amounts of data until it establishes the rules to drive. Of course, not every driverless car on the road is trained from the bottom up like this. Programmers take a base model, teach that network from the start, then clone and place it in an entire fleet of vehicles.¹³ 

Fully autonomous self-driving cars are not commercially available; it will still be some time before that becomes a reality.  Until then, vehicles as we know them will gradually become more automatic, evolving until they can drive independently. Before becoming fully autonomous, they will share driving responsibilities with us as we work together to navigate the roadways successfully. As the self-driving function increases, driver responsibility will decrease until we are no longer required to assist. In 2014, the US Society of Automotive Engineers developed a classification system outlining the progression levels of self-driving automation within vehicles, ranging from zero to full automation.¹⁴ 

Self-Driving Levels

• Level 0 – No Automation- The traditional vehicle where we drive.
• Level 1 – Driver Assistance- The vehicle has limited built-in automation, like adaptive cruise control or lane-keeping, but we still do the driving.
• Level 2 – Partial Automation – The vehicle features multiple types of automation that assist with steering, speed control, lane keeping, and maintaining a safe distance. The car can drive for very short durations, but the driver is in control and closely supervises.
• Level 3 – Conditional Automation – The car has a brain and can drive autonomously for longer durations in certain situations, such as on highways. The vehicle can be left to drive independently within its designed conditions, but we must be ready to step in at any moment if it gets stuck. 
• Level 4 – High Automation – The vehicle does all the driving under normal conditions when inside the footprint of a predefined zone, allowing us to sleep in the backseat if we want. We would only have to step in and assist when the car is outside its designed footprint or if conditions such as weather become too challenging. Even if the vehicle got stuck, it would still respond by pulling over safely to pass control over to us. With this setup, the car is the primary driver, doing most of the driving. We are the secondary driver acting as a supervisor who steps in when the vehicle is stuck or if we just want to be the one driving.
• Level 5 – Full Automation: No driver, pedals, or steering wheel are required; the vehicle does 100% of the driving at all times. ^{15 16}

The highest level of commercially available self-driving vehicles is level 3 conditional automation.¹⁷ This design allows the car to drive independently for hours without needing assistance, such as on highways. Still, at any moment, it could encounter a scenario where it needs to tag the driver to take over.¹⁸ When the self-driving system is driving, we are the supervisor, watching it and taking the wheel if it encounters a problem it cannot handle. The two systems coalesce in a single vehicle, leveraging a pseudo design where they work as one unit to navigate the roadway safely.¹⁹ 

A self-driving network trains to handle most situations, but there will be times when it encounters something new for which it has no answer. When that happens, the vehicle will disengage and pass control over to the human driver.²⁰ We take the wheel, acting as a secondary processor that can draw on a lifetime of experience to bail out the self-driving system. ²¹ 

Disengagements range from minor to severe and serve as new lessons that the network learns to make it an even better driver in the future. The disengagements could be due to overhanging branches appearing as obstacles, confusion with vehicles in other lanes, or not seeing another car leave a garage.²² When these events occur, the human driver takes the wheel to get out of harm’s way and returns control to the vehicle once the obstacle is clear. The network then updates its model to have an answer for the next time it encounters a similar situation. That update may not be in real-time, but the general idea is that disengagements continually tune the network’s model, improving its driving ability. On top of that, the lessons learned from each vehicle are shared collectively across the entire fleet, allowing them all to learn from a single disengagement.²³ 

 As of 2018, Google’s driverless car company Waymo, has vehicles travelling 11,000 miles between disengagements.²⁴ On the other hand, humans drive hundreds of thousands of miles without incident.²⁵ Self-driving cars and humans will continue to work together for a while before vehicles can drive independently. Future chapters will continue to develop the idea of two systems working as one to navigate the environment, as it applies to being a human with a brain. 

Endnotes

1. Wadhwa, Vivek, and Alex Salkever. P.149. The Driver in the Driverless Car: How Your Technology Choices Create the Future. Berrett-Koehler Publishers, 2019.
2. Wooldridge, Michael. P.176. A Brief History of Artificial Intelligence: What It Is, Where We Are, and Where We Are Going. Flatiron Books, 2021.
3. Burns, Lawrence D., and Christopher Shulgan.P.224. Autonomy: The Quest to Build the Driverless Car—And How It Will Reshape Our World. Ecco, 2018.
4. Shane, Janelle. P.55. You Look Like a Thing and I Love You: How Artificial Intelligence Works and Why It’s Making the World a Weirder Place. Voracious, 2019.
5. Wadhwa, Vivek, and Alex Salkever. P.106. The Driver in the Driverless Car
6. Dormehl, Luke. P.172. Thinking Machines: The Quest for Artificial Intelligence–and Where It’s Taking Us Next. TarcherPerigee, 2017.
7. Burns, Lawrence D., and Christopher Shulgan.P.224. Autonomy
8. Du Sautoy, Marcus. P.84. The Creativity Code: Art and Innovation in the Age of AI. Belknap Press, 2019. Kindle file.
9. Shane, Janelle. P.9. You Look Like a Thing and I Love You
10. Shane, Janelle. P.115. You Look Like a Thing and I Love You
11. Goleman, Daniel. P.164. Focus: The Hidden Driver of Excellence. Harper, 2013.
12. Du Sautoy, Marcus. P.188. The Creativity Code
13. Hawkins, Jeff. P.182. A Thousand Brains: A New Theory of Intelligence. Basic Books, 2021.
14. Wooldridge, Michael. P.149. A Brief History of Artificial Intelligence:
15. Shane, Janelle.P.59. You Look Like a Thing and I Love You
16. Wooldridge, Michael.P.155. A Brief History of Artificial Intelligence
17. Shane, Janelle.P.59. You Look Like a Thing and I Love You
18. Shane, Janelle.P.59. You Look Like a Thing and I Love You
19. Shane, Janelle.P.20. You Look Like a Thing and I Love You
20. Shane, Janelle. P.59. You Look Like a Thing and I Love You
21. Wooldridge, Michael. P.155. A Brief History of Artificial Intelligence
22. Shane, Janelle. P.57. You Look Like a Thing and I Love You
23. Topol, Eric. P.87 Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again. Basic Books, 2019.
24. Wooldridge, Michael. P.153. A Brief History of Artificial Intelligence
25. Wooldridge, Michael. P.154. A Brief History of Artificial Intelligence