The Gap in Autonomous Self-Driving Vehicle Development, Part 1
As recently as 4 years ago, many major auto manufacturers declared that they would have self-driving cars on the road by 2020. Now, industry experts are saying that we are still over 10 years from being able to go out and buy a self-driving vehicle. So what went wrong? It is not for lack of trying. According to the website The Information (subscription required) “A group of 30 companies has spent at least $16 billion on developing fully self-driving cars over the past few years — and so far they have little, if any, revenue to show for it. But billions more will likely be needed before the technology is ready for primetime.” There are many smart people working on “solving” autonomous driving, and companies are clearly willing to throw plenty of money at the problem, but there is still a large gap to achieve the result we want. In this 3 part essay series I will explain why and propose a smarter way of making self-driving vehicles a reality. I will look at 3 areas where investment, ingenuity, innovation and collaboration should be focused.
Part 1: The Unpredictability Problem
“Self-driving cars rely on artificial intelligence to work. And the 2010s were a great decade for AI. We saw big advances in translation, speech generation, computer vision and object recognition.” (Vox)
The current approach to self-driving vehicles is to try and replace the operator with cameras, sensors and a computer with artificial intelligence (AI) software acting as the “brain.” It is easy to see where the companies are spending their money, the vehicles they are testing are veritable rolling laboratories outfitted with radar, LIDAR, ultrasonic sensors, cameras, vehicle dynamic sensors, accelerometers and GPS chips for location. They also need devices to control steering inputs, acceleration and braking. On top of all that, they need a computer fast enough to run the AI software. The AI must compile and analyze all the data it receives (up to 4TB of data per day per car), determine a response, then send the commands to the vehicle control system to execute the response; and it must do this many times a second, while the vehicle is traveling between 30 and 80 miles per hour. This is a lot to ask.
These companies all have armies of engineers, computer scientists and researchers working to develop the software and “teach” the AI how to be a good driver. While the advances in artificial intelligence are exciting, according to J.D. Power “industry experts say that perfecting self-driving technology is proving more challenging than originally thought.” The reason why it is so challenging is also the reason why a large gap still exists in self-driving vehicle development. The AI is being asked to predict what will happen in a very unpredictable environment. Currently these companies are trying to collect as much real world data as possible to teach their AI how to categorize different variables, and combinations of variables, then how to react to each scenario. In lieu of real world data, the companies are feeding their AIs simulated driving data to try and speed up the process. Per Vox, “Even with extraordinary amounts of time, money, and effort invested, no team [can] figure out how to have AI solve a real-world problem: navigating our roads with the high degree of reliability needed.”
The difficulty of such an approach is essentially like trying to force a square peg into a round hole. All of these AI controlled, self-driving laboratories are like a super shiny, expensive square peg that doesn’t fit into the round hole that is our current road and highway infrastructure. Let’s look at a real world problem and do some math to understand the size of this problem. The image below (figure 1) shows an intersection in New York City. We can categorize and roughly count the different variables.
People: 33, Cars: 18, Bicycles: 2, Traffic lights: 2
Now let’s calculate how many distinct scenarios can occur based on the variables above. The following equation is used to calculate the number of unique combinations we can create based on a population size, n, and the sample size, r.
Based on the variables we counted above, n will equal 55, the total number of variables. Now let’s assume that 10 of those 55 variables will do something unpredictable at a given moment, which makes r equal to 10.
That is over 29 billion different combinations or potential scenarios that a self-driving vehicle will need to assess and react to, all within fractions of a second. Self-driving vehicles work great in controlled environments, but when you increase the number of variables, the environment becomes less predictable and they struggle. If these companies continue to throw money and engineers at the problem they can eventually shave down their square peg and make it round, but, per the experts, that will likely take 10+ years and many more billions of dollars.
In Part 2 of this essay series will look at how we can lower the both the n and r values discussed above, and make the round hole more square.
