Frank Chen of a16z on 16 Things About Autonomous Vehicles

In a recent a16z presentation, Frank Chen, a partner at Andreessen-Horowitz, says that everything that moves will eventually go autonomous. To be honest, a lot of that was clear at CES recently, where autonomous cars and drones were ubiquitous. If there is a reason for an object to move by itself, it eventually will be.

Frank runs their Deal and Research team, so he sees all of the entrepreneurs who walk in the door and ask for funding. They see lots of AI startups, but he focuses on cars because it is such a huge market. He says there are sixteen questions—some technical, some business, some social implications—that are asked all the time at his investment firm, and he presented these answers. I think he picked 16 because Andreesen-Horowitz is often called a16z (because there are 16 letters between the a and the z).

How  will we get from SAE level 0 (or 1 where most cars are today) to level 5 (no steering wheel)

• • If you are an incumbent car manufacturer, chances are, you will take incremental steps until you reach full autonomy in the driving experience.
  • However, if you are Google (or someone like Google), then you could enter the market with a fully autonomous vehicle. Since one of the biggest problems of autonomous driving is managing the user experience when switching between driving and self-driving modes, it may be easier to enter the market at level five.

Will cars have Lidar?

• • Lidar, which stands for Light Detection and Ranging, is a remote sensing method that uses light in the form of a pulsed laser to survey the environment. They are those ugly spinning things on the roof of some Google's cars and cost $75K today. So for sure not that sort. But Lidar is becoming solid state and once they cost about $250 it will probably be used, although right now Tesla is trying without.

What new types of pre-computed HD maps will be necessary for the autonomous car, or will the maps be built on the fly?

• • Google map-type technology is not yet good enough to use for autonomous driving since they are built for humans, not computers. There is not nearly enough resolution in what we have now. (Plus, who will provide the maps, since there are only three mapping companies left?)
  • We could figure it out with a supercomputer in the trunk, but the power budget may impact the range too much, among other challenges to consider.
  • For sure a combination.

What will the blend of software techniques be?

• • The red-hot area is end-to-end deep learning, where you put the sensors on one end and actuators on the other, and the system “learns” quickly how to react in certain scenarios. But there are areas where there is probably not enough data to learn that will require algorithms (such as what to do when an ambulance comes along). This would rely on other sources of computing.

How much real-world versus virtual-world testing can you expect? Can you avoid doing the millions of miles of driving required to “teach” an autonomous vehicle how to drive?

• • That remains to be determined. Almost certainly a blend. We will have to see over time.
  • It all depends on the sophistication of the virtual world. The answer to the question is to be determined.

How will vehicle communication be implemented? Will V2X radios play a major part?

• • V2X is a vehicular communication system that incorporates other more specific types of communication, such as V2I (vehicle-to-infrastructure), V2V (vehicle-to-vehicle), V2P (vehicle-to-pedestrian), V2D (vehicle-to-device) and V2G (vehicle-to-grid).
  • Mercedes is putting V2V and V2X in all S-classes in the next year or two, but that just shows you the deployment problem: it only works on other Mercedes S-series cars!
  • Also, who decides on the protocol compatibility of divergent devices?
  • It’s likely not going to be widely deployed all the way anytime soon.

When will we be able to take out traffic lights and 4-way stops?

• • Eventually. But all the V2X stuff needs to be ubiquitous.

How will automakers localize their cars to learn the local driving style?

• • Probably via social learning. Ship the car, and it learns the local driving conventions, rather than building special cars for each market.

Who will win: Silicon Valley, China, or the incumbents?

• • Obvious bet is the incumbents. They all have offices in Silicon Valley because they know the computing power that will be required to get there.
  • It could also appear as a new company in Silicon Valley; it’s been known to happen.
  • Don’t rule out Chinese manufacturers; they have published more deep learning papers than anywhere else on the planet.

Will we continue to buy cars or transportation-as-a-service?

• • If we buy transportation as a service (like from Zipcar or Uber or Lyft), the industry will become more like the airline biz. Car companies will become b2b companies, like Boeing and Airbus are.
  • Be prepared for drastic changes in the transportation industry.

How will insurance change?

• • Your demographics (sex, age, etc.) and cost of the vehicle will no longer matter when determining your rates.
  • What may matter is the safety record of different manufacturers—the safest companies will get the best rates.

How will accident rates change?

• • Currently, accidents are caused primarily by user error.
  • Accident rates will eventually go down. Most accidents with Google cars are because people expect the car to do something and it doesn't. Google cars don’t drive like human drivers.
  • Even if the accident rate goes up a little bit at first, the learning curve will be better for cars than humans.

Will it become illegal to drive? When?

• • It should happen as fast as possible! Human drivers are fallible, much more than robots.
  • Driving can be fun for some, but that fun shouldn’t happen on the road, where people’s lives can be lost to the fallibility of human drivers. Driving can become something like driving race cars on a track.

How will commute times change?

• • It’s hard to predict! On one hand, having someone else doing the driving allows us to focus on other things, so an increase of driving times might not matter to the passengers, and thus it could go up.
  • On the other hand, with V2X communication, it will become easier to navigate through hazards and restrictions, such as closed lanes or weather conditions, so the traffic might ease up.
  • The answer to this question is unclear.

How will cities change?

• • Autonomous driving can free up space in cities that is currently used for our current infrastructure, such as parking lots, car repair, filling stations...
  • When cars got cheap, it was easy to predict we would all own them. But, like many couldn’t predict the ultimate “cost” of Wal-marts offering cheap products appearing all over the country, there are bound to be unintended consequences and results of changes in our transportation network.

When will this happen?

• • Predictions vary from 2018 to 2040. Tesla predicts 2023. Uber says their fleet will be autonomous by 2030. IEEE says 40% of all cars on the road will be autonomous by 2040.
  • Regardless of who is right, chances are that it will happen in our lifetime—something Frank wouldn’t have thought possible, ten years ago!

 Watch the video (about 30 minutes):