中文 English

Self-Driving Cars In San Francisco

Autonomous driving systems are rapidly gaining proficiency, but there’s still areas where they don’t scale.


You probably have heard that Waymo has completely driverless (no safety driver) taxis serving Phoenix. 600 of them. But you can’t go and buy one. Why is that?

Paul Graham, the founder of the incubator Y Combinator, is celebrated for many reasons, but two things he has said have become mantra in the startup world:

Build something people want.
Do things that don’t scale.

When it comes to autonomous cars, I think it is pretty obvious that this is something people want. As an aside, I think a lot of EDA startups have failed over the years because they have built a technology that was difficult and fun to build…but it turned out nobody wanted, at least not yet. See my post It’s the Second Mouse That Gets the Cheese for more on that.

So what about doing things that don’t scale? Isn’t that totally wrong? As Paul Graham put it, though:

A good metaphor would be the cranks that car engines had before they got electric starters. Once the engine was going, it would keep going, but there was a separate and laborious process to get it going.

Isn’t a company like Airbnb all about working at scale, running everything through automated smartphone interfaces so that people can sign up to rent their place, sign up for a rental, pay, and so forth? That’s true…today. But that’s not how it started. Paul again:

Airbnb is a classic example of this technique. Marketplaces are so hard to get rolling that you should expect to take heroic measures at first. In Airbnb’s case, these consisted of going door to door in New York, recruiting new users and helping existing ones improve their listings.

Airbnb now seems like an unstoppable juggernaut, but early on it was so fragile that about 30 days of going out and engaging in person with users made the difference between success and failure.

The Cruise video

I ran across this video recently. It is Cruise CTO Kyle Vogt with OpenAI CEO Sam Altman. Coincidentally, Sam took over as CEO of Y Combinator when Paul moved on to do other things. It is a 75-minute drive through San Francisco, showing the current state of the art. But don’t worry, you don’t have to commit to over an hour since the drive is sped up by a factor of 5X. The video is 25 minutes long, and if you want to see the state of the art in autonomy today, I recommend watching it. Don’t be too distracted by the lower part of the video, showing the street. At least sometimes, look at the top half of the screen showing what the car is recognizing (cars, pedestrians, bicycles, the roadway, and so on. As Sam Altman says at one point in the video, “I can’t tell that this is not being driven by a human driver.”

Things I think are especially interesting to watch:

The Cruise system doesn’t just identify other vehicles, it predicts what they are going to do.

  • 2:29 Unprotected left turn (at lights, but with no green arrow).
  • 3:14 Goes around construction that partially blocks the street. This required “remote assistance” or “phone a friend” as Kyle called it. Happens “every 5 or 10 miles right now.”
  • 4:56 Kyle says they have driven “a couple of million miles in San Francisco right now.”
  • 5:10 Simulated pickup/dropoff at the curb.
  • 5:25 Slows down due to danger from cars parked at 90° angle and an insane driver passes it at speed.
  • 8:30 Very smooth lane change.
  • 9:20 Picking the lane with the least traffic.
  • 9:45 Going around a double-parked car.
  • 11:15 Going around a parked truck when there is oncoming traffic, too.
  • 13:00 “There are between 1 and 2 dozen neural networks.”
  • 13:35 “Each car we are predicting if it is a traffic participant or something we have to go around.”
  • 15:00 Avoids cyclist going the wrong way.
  • 16:15 “This version is designed to work most of the time in San Francisco…if there is too much fog or rain it’ll kick itself out and pull over.”
  • 17:20 Bike weaving between the lanes.
  • 18:15 Trajectory of improvement is “3X to 10X per year. There’s no reason it will top out at human performance, it’s just going to keep going at 3X to 10X per year.”
  • 18:55 “The takeaway I have from this video is that it is already superhuman in some aspects.”
  • 20:55 “I think we are badly underestimating what it means to compound at this rate for ten years.”
  • 23:40 “We’re not even talking about the multi-agent case…then you can do really interesting things like where the robots drive within 6″ of each other at speed on existing roadways.”
  • 25:40 Ends.

Maturing autonomous vehicles

It goes without saying that autonomous vehicles have not reached the same level of maturity as Airbnb. They are still at the stage of doing things that don’t scale.

To my mind, the two things that don’t scale yet are cost and occasional manual intervention.

The cost of the electronics, and especially the lidar (except for Tesla which doesn’t use it) is probably way too high. Tesla proved that there is a market for a very high-priced entry into the market since their first offering was a two-seater roadster priced at around $100K sold more as a rich plaything. So in a sense, since there is a very limited market for $100K cars, this was something that didn’t scale. The strategy today for someone like Cruise is to get the technology to be great and then work at cost-reducing it. Of course, lidar suppliers like Velodyne are working on that part of the problem.

There is still a need for occasional manual intervention, as in the Cruise video. This scales to a few hundred taxis requiring occasional intervention. It does not scale to millions of cars. Of course, in some cases the person in the car is available. It is not clear whether “driver” or “passenger” is the best term for that person, they are blurring. But fully autonomous vehicles, even yours, may not always have you in it: it is coming or going from a parking lot, or driving to your kids’ school empty to pick them up. Or driving back with them in it, so although the vehicle is occupied there is nobody in it who can help in a “call a friend” situation.

Obviously, two things have to happen to make this more scalable: the cost has to come down, and manual intervention needs to drop to close to zero. But the basic driving technology is getting tantalizingly close.

I’ve tried it

I’ve ridden in a driverless taxi in Las Vegas during DAC 2019. The vehicle was an Aptiv BMW. Unfortunately, I was not allowed to take any pictures inside the vehicle, just outside as you can see.

The main display was similar to the top half of the cruise video. There is no need to show the video from the front of the vehicle since you can just look out of the windshield. The vehicle did have a safety driver, unlike those Waymo taxis in Phoenix. I don’t know if the Cruise vehicle in the video had a safety driver. It had to get offline help to get around construction, as you saw if you watched the video, but that was not from inside the vehicle.

Leave a Reply

(Note: This name will be displayed publicly)