Original article was published on Artificial Intelligence on Medium
Self-driving cars are the case study we focus on because it’s what we spend most of our time in (and there’s been more than 270 million registered vehicles in the US alone in 2018). The story of self-driving cars crossed with engineering started with the DARPA Grand Challenge in 2004. A 150-mile long course through the desert.
In 2004, 0 cars made it 10 miles.
In 2005, a couple finished the course. The video below shows you how different the view of a self-driving car was then from now. This was pushing the limits of autonomy.
Now, self-driving cars are a much more defined area of research and development. This isn’t an article about where self-driving cars are at a technological level. The government even has websites now defining that.
This is an article that focuses on one patch on one problem.
The problem — Corner cases: self-driving cars are bad at dealing with unforeseen environments like construction, fallen trees, road damage, etc.
The solution — Remote control: call a human in to figure out how to maneuver around the rare obstacle.
Why this works: with enough sensors and computers, we can make self-driving cars incredibly safe. When they aren’t sure of something, they go slow. When they’ve never seen something before, they stop.
A stopped car isn’t very useful though. That’s when we have human overseers call in, look at the video feed, and drive around the challenge. Everyone wants their Tesla to be fully autonomous, but we need some way to get data for the .0001% of miles that the car hasn’t seen, or may never see.
Teleoperation of autonomous cars
There’ll be office buildings for Uber drivers (or maybe they’ll be outsourced to India). Here, the drivers will be sent mini-driving challenges. The mini-challenges will be hard for a computer, but they’ll be trivial for a human. Something like: there’s a man with a slow sign and the road is now one way, or: there’s a trash can in the road. The human operator needs to assign 3 waypoints around the obstacle.
In a testing fleet, say a car is stuck at a stop sign or unknown obstacle 10% of a time. That means 1 human could cover 10 cars ideally, but say 5 cars to be safe. That’s 1/5 the driver cost that is making Uber burn cash.
This factor scales incredibly well. Consider a beta round of autonomous vehicles with only 1% unknown vehicles. Now 1 driver controls 50 cars. This scaling will only improve with time. It results in a dramatic reduction in costs.