You are here

Aurora's manifesto is worth a read

Aurora test vehicles

I rarely just link to another story, but today I will point all to Aurora's post on its development philosophy. I think it contains some important lessons for all in the field.

It won't surprise readers to know I agree with pretty much this whole document, having said many of these things myself and also having learned some of them from or along with Aurora CEO Chris Urmson while at Chauffeur (Google/Waymo).

Aurora has already had stellar success. They have been funded to a high level, have done partnerships with several major OEMs, and even turned down "unicorn" level acquisition offers, something ridiculous for such a young company.

The path to this is now clear. Waymo has a serious lead over everybody, including Aurora. And the other players are learning that, but they know that Waymo isn't particularly keen to partner with them in any way they would like, preferring to take a shot at the brass ring -- as they should.

This leaves the larger players very interested in partnering with the man who ran Waymo's engineering efforts (and several other credits.) It's the clearest path to catching up. And Aurora is happy to be a partner and deal with any good partner, to be the neutral Switzerland and not threaten to compete head to head. (At least for now.)

All the big players know that they want to be the party to own the customer. The one who sells you the ride, or the whole car. The one who gets the money and then decides how to distribute it to all the suppliers. It's the seat you want to have. But being the best supplier of secret sauce who is willing to deal is a pretty good seat -- a fantastic seat for a brand new company with no skin yet in the game.

Cruise, if spun off from GM, might try for that seat as well. So will a few Tier 1 suppliers. Uber and Lyft may even offer their software stack for free but with a big catch -- you must exclusively be in their ride network, and let them own the customer.

Strap in, it's going to be a bumpy ride.

Aurora's advice

Some key elements of the advice are below, and I've written articles on many of these concepts, but you should read the original post

  • This is the modern version of rocket science, it needs exceptional skill
  • The idea of incremental "levels" is broken, but you can still develop incrementally by going after bigger and bigger segments of the grand problem of driving everywhere.
  • Test heavily in sim before the road. (This is not something Waymo or anybody did extensively in the earliest days, but times have changed.)
  • Go for the big, revolutionary approach and prototype it quickly. Don't try to build a ladder to the moon.
  • Be measured in how you use machine learning. Rely on existing tools and give them the best (not the most) training data. Design to use it from the start.
  • It's foolish today to try and do it all with one, or even two sensors. Get all the independent information you can (LIDAR, radar, vision) to get as safe as you can, quickly. Later try to make it cheap by cutting out sensors.
  • Maps!
  • Combine machine learning with the best of robotics engineering; don't imagine it will work end-to-end.


The use of high definition maps to spectacularly reduce the complexity of the task is a great example of finding a way to constrain the problem to reduce the reliance on AI.
My experience with deep learning is very limited, but from what I have read it is not clear that they have the power to be much more than classifiers.
I am making the potentially unwarranted assumption that the sort of understanding of the real world required for driving is not anything like human level as even simpler animals are able to perform object recognition, navigation and avoidance.
But I am not sure that even if we hade sufficient processing power that Artificial Neural Nets would be up to the task.
I feel like we are waiting either for a much deeper understanding of why animal brains, even small ones, can do what they do, or another way to constrain the problem before we go from 99.9% to 99.99%

Large neural networks are very impressive. They have adherents who claim that there is no obvious limit on their capability, if they get large enough. They wonder if natural brains are a great deal more than neural networks. Of course that is not a universal holding, but people do indeed try to drive a car by taking pixels into a neural network and getting paths out. I, and Aurora, agree that this is not very likely any time soon, but others feel it is a worthwhile bet to try.

On the other hand, neural networks do have productive use in things like choosing smooth and natural short term paths, and other tactical things. And of course, classifier, including things like identifying the intent of other road users, all the way to examining human posture and facial expressions.

I found it disappointing that Chris wrote about their engineering culture without making it clear to his future employees how they must integrate their technology into society. He talks about "safety", but you get the feeling it's "we have to introduce this safely otherwise we'll pull and Uber and get shut down".

I would prefer riding in a self driving vehicle designed by a team humble enough to say, "we are bringing something to the table, but it is our responsibility to find out how we can fit this into society, and we will actively look for where our new thing is dumping risk, wasted time, fear, etc onto other users of the transportation system".

That is a topic that does get talked about. Every talk I've ever seen or given, including Chris' talks, has indeed put the focus on the safety goal, and it is a real goal. I agree there's not a lot of focus in every talk about social integration, but it certain gets covered. This is a technical manifesto so it's mostly about how to make it work, which is almost entirely how to make it safe.

The safety goal is not as cynical as you imagine. I know these folks and worked with many of them at the dawn of this, and while there is no question that many enjoy working on the technology because it is incredibly cool technology, they are also quite attracted by the opportunity to make a real change in the world by improving road safety. Many of them because they have had personal losses due to car accidents. This is actually more of a motivator than the motivation for huge financial gain. So I don't think you would find many who are working on safety because they don't want to "pull an uber." They want to improve safety.

Add new comment

Subscribe to Comments for "Aurora's manifesto is worth a read"