Get best of the week

Starsky Robotics unmanned trucks come to an end

image

March 19, 2020

In 2015, I became interested in the idea of ​​creating unmanned trucks and founded Starsky Robotics. In 2016, we released the first fully unmanned vehicle allowed on the road, which carried passengers for money. In 2018, we released the first unmanned truck allowed on the road, which made a full-fledged raid, albeit on a closed road. In 2019, our truck became the first fully unmanned freight vehicle to drive along a busy highway.

And in 2020 we are closing.

I am still incredibly proud of the product, team and organization that we were able to create. I am proud of the organization in which doctors of science and truck drivers worked side by side, where the problems of generations were solved by people whose minds exceeded their regalia, and where we learned how logistics will work in the future.

Like Shackleton on his expedition to Antarctica, we did what no one else did. However, like him, things didn’t go according to plan.

So what happened?

image

Most of the Starsky team in February 2019. Nothing in my life made me so proud as working in this incredible team.

Time, I think, is precisely guilty of our sad fate. I believe that our approach was correct, but too many unfulfilled promises in the field of AI have accumulated that have not yet found their practical solution. Since these breakthroughs did not appear, the downpour of investor interest turned into drizzling rain. It also did not help that last year’s high-tech IPOs took away a lot of energy from the technology industry, and freight transport has been declining for about 18 months.

Unmanned Vehicle Area


There are too many problems in the field of unmanned transport, which does not allow us to give a detailed description here: the professorial pace at which most teams work, the lack of tangible milestones in deployment, the well-known secret about the non-existence of the Robotaxi business model, etc. The biggest problem, however, is that controlled machine learning does not justify the hype. This is not real C-3PO-like artificial intelligence, it is a sophisticated pattern matching tool.

We will return in 2015, when everyone thought that their children would not have to learn to drive a car. Controlled machine learning (under the name of AI) developed very quickly - in just a few years, it went from recognizing cats to more or less acceptable driving. AI seemed to follow Moore's law:

image

According to forecasts, all of humanity will certainly be economically uncompetitive in the near future. We will need a guaranteed income, which will help to cope with the emerging gap between the machines and us.

Five years have passed, and experts in the field of unmanned vehicles no longer promise the creation of a full AI after the next commit to the source code. On the contrary, the situation is now such that we have moved away from creating unmanned vehicles for at least 10 years.

It is widely believed that the most difficult part of building AI is the formation and adjustment of its behavior in situations that occur infrequently, i.e. in borderline situations. In fact, the better your model, the harder it is to find reliable data sets about such situations. In addition, the better your model, the higher the accuracy requirements for the data needed to improve it. Instead of observing an exponential improvement in AI performance (Moore's Law), we see an exponential increase in the cost of improving AI systems — controlled machine learning seems to follow the S-curve.

image

We are talking about the S-curve for the reason that Comma.ai, with 5-15 engineers, produces a performance that does not differ much from the successes of Tesla, in whose team more than 100 people. And also because we at Starsky managed to become one of the three companies conducting tests of unmanned vehicles on public roads (with the efforts of only 30 engineers).

This situation is not unprecedented - S-curves are often applied in the field of technology (Moore’s law actually consists of a number of S-curves, since chip technologies continuously replace each other, and the best in this race increase the overall curvature of the graph). The problem is trying to compare the capabilities of modern technology with how well people can drive. I would suggest that there are possible options: we have already surpassed the human level (mark L1), we almost managed to do it (L2), we are very far from that (L3).

image

If L1 is the level of human driving, then the leading companies in the field of unmanned vehicles simply have to prove the safety of their decisions, and they can deploy them. I do not think anyone believes in this, but it is possible. If the level of human driving is at around L2, then large teams will need from 1 to 25 billion dollars to solve all the problems. When large investors in the unmanned vehicle market say that this industry is suitable only for large companies, they are voicing this particular rate. And they do it. If the level of human equivalent is at L3, then hardly any of the existing technologies will make this leap. Whenever someone says that 10 years are left before the creation of unmanned vehicles, these are almost certainly true considerations.Not many startups can survive 10 years without releasing their product. All of this means that almost no modern team will ever release AI systems that make decisions.

Not many startups can survive 10 years without releasing their product.

Why we did not survive


For a person who is not familiar with the dynamics of attracting venture capital investments, all of the above may seem to be an excellent occasion for investing in Starsky. We did not have to create a “real AI” to be a good business (we thought it could cost about ~ $ 600 / car / year), so we expected that we could raise funds, despite the fact that the above is becoming more and more obvious. Unfortunately, when investors cool to any area, they tend to cool to the whole market. We also saw that investors really did not like the operator’s business model, and that our heavy security investments were not transferred to investors.

Truckers Blues


If the teleoperator solves half of the problem of unmanned transportation, the second half is solved by the operator. The carrier can choose a deployment location, which will make decisions about working methods. Your system only needs to be safe on the selected routes and in the selected conditions (that is, work reliably on the easiest routes, and in bad conditions - stop and wait).

The specifics of the freight market participants also influence the decision to be an operator. Freight companies are not very technological customers (see what they use), and none of them understands the purchase of road robots with high safety requirements. Even if Starsky had improved unmanned technology and worked well in terms of security, it would have taken years to deploy enough systems to make the right profit.

image

“You can always understand how seriously the company takes unmanned technologies, by how seriously they take teleoperators,” the supplier once told me. Nevertheless, we found incredible resistance of industry and investors to our approach based on the use of a teleoperator.

Despite the fact that freight carriers do not understand anything in the procurement of robots with increased safety requirements, they understand how to buy freight transportation capacities. Each large shipping company works as follows: their brokers buy capacities from smaller transport companies and operator owners, many of which they do not allow too close, because they do not know how much you can trust the safety indicators from their own reports. In Starsky, we found more than 25 brokers and carriers who are ready to ship goods on unmanned trucks. Although this is a business with a lower profit margin (compared to the traditional software market with a 90% profit margin), we expected to be able to achieve a 50% margin on time.

It took me too much time to realize that venture investors would prefer to invest 1 billion and get a 90% margin, instead of investing $ 5 billion with a 50% margin - even if the capital requirements and growth were the same.

And the growth would be the same. The biggest obstacle to deploying unmanned vehicles is not sales, but safety.

Nobody likes to invest in security, everyone likes the functionality


In January 2019, we met with the head of the security department and the head of the public relations department in the conference room for a strategic session. The question was how to make security look attractive enough to fully ensure it. The month before, we publicly released VSSA, a white paper that details our approach to security. We passed it on to the most sensible journalist, but instead of covering it in detail, the media mainly wrote about teleoperations. We left the meeting in a nervous state - we could not figure out how to make safety engineering attractive enough to start writing about it.

And we never came up with a solution.

image

Ironically, we planned to launch a fleet of 10 v2 trucks by January 2020. These systems were designed in such a way that we could demonstrate the safety of the entire fleet, which would ensure the regular operation of unmanned services by June 2020.

The problem is that people get excited by things that rarely happen. An example is Starsky's unmanned vehicle road test. Even when it is negative, much more than 100 people who die every day in car accidents report a plane crash. Work on security by definition is the creation of a system that works without exception.

Safety engineering is the process of thoroughly documenting your product, which will allow you to know exactly the conditions under which it will fail, as well as be able to assess the severity of these failures and measure the frequency of occurrence of these conditions. All this will allow you to understand the likelihood that your product will harm people, compared with what damage you consider acceptable.

It is very, very difficult. It is so complicated that from September 2017 until our unmanned trip in June 2019, we did just that. We documented our system, built a secure backup system, and then repeatedly tested our system for failure, corrected these failures and repeated.

image

The problem is that all this work is invisible. Investors expect the founders to lie to them - how can they believe that the unmanned run we did actually only have a 1 in 1 million chance of death? If they don’t know how difficult it is to make an unmanned run, how can they know that someone else will not be able to do it next week?

Our competitors, on the other hand, have invested their engineering efforts in creating additional AI capabilities. An example is decision-making systems that can sometimes change lanes, or can drive along aboveground streets (provided that they have enough map data). This is really cool, advanced technology.

Investors were impressed. It didn’t matter that the transition from “sometimes working” to statistical reliability required 10-1000 times more work.

image

So what's next?


Around November 12, 2019, Series B collapsed, our $ 20 million project. We sent most of the team on the 15th on unpaid leave (probably the worst day of my life), and then we started selling the company and made sure that not a single member of the team would be left without shelter (or a visa or health care for new and future parents).

By the end of January, we were able to employ many unprotected workers, and now I am selling the company's assets (including a number of patents necessary for the operation of unmanned vehicles). As the captain of a sinking ship, I put most of the crew on lifeboats and now see how icy waters are approaching my feet, while I begin to think about what to do next.

From my point of view, the most probable line of the level of human driving is L3, which means that no one should build a business on the creation of safe AI decision-making systems. Current companies will continue to lose momentum over the next two years, followed by several years with virtually no investment, and (hopefully) another test of unmanned vehicles on the highway for 5 years.

I would like to make a mistake. An aging workforce that will almost certainly begin to limit economic growth in the next 5-10 years; 4,000 people who die annually in car accidents, seeming an unnecessary sacrifice. If we showed something in Starsky, then this is something that can be achieved if we sincerely focus on eliminating the need for a person to be at the wheel in certain cases. But you will need someone who can bring this idea to life.

Goodbye
Stefan.




image

About ITELMA
- automotive . 2500 , 650 .

, , . ( 30, ), -, -, - (DSP-) .

, . , , , . , automotive. , , .

Read more useful articles:

More articles:


All Articles