BETA
This is a BETA experience. You may opt-out by clicking here

More From Forbes

Edit Story

Starsky Robotics Shuts Down And Worries Everybody Else Will Also Fail In Robotic Trucks

Following
This article is more than 3 years old.

In March, Stefan Seltz-Axmacher, CEO and founder of Starsky Robotics, announced the end of their trucking company, which combined autonomous driving software and remote monitoring techniques to move big rigs. Back in 2015, I had advised Seltz-Axmacher on their plan, but had lost touch. Recently he published a detailed essay outlining the causes of the failure of the company which caused some stir, because one of its assertions that effectively, the problem is too hard, and others will fail as well. Others have rejected that claim and argue that Starsky failed because of failings of the company itself, not the industry.

Let’s look at some of the arguments made in the essay:

  1. Supervised machine learning just isn’t up to the autonomous driving task, and may be very far from getting there.
  2. While there should be low-hanging-fruit in the trucking world for robotrucks, the industry is hidebound and not ready.
  3. VCs don’t like to invest in businesses which must include a traditional, lower-margin, more capital intensive component, such as owning and operating a freight delivery service.
  4. Everybody talks about how Safety is #1, but it really isn’t and doesn’t attract attention or investment.

I’ll actually start with some elements of #2 that I dug into in a discussion last week with Seltz-Axmacher. Starsky picked trucking because it’s a horribly managed industry with lots of opportunities to do better. There’s a giant driver shortage. Drivers are also sometimes unreliable — as one customer told him, a robot is unlikely to get into a fight with people at a warehouse, or decide to stop in Las Vegas and spend a few days in a strip club. This doesn’t happen all the time in trucking, but it happens enough that having it never happen is a big plus. Starsky planned to run their trucks only on the easiest of highways, short to middle distance runs on clear, open uncongested highways. If a problem arose due to traffic or weather, the truck would just pull over — because multi-hour stops are common with human drivers too.

This also meant that a combination of full self-driving with occasional remote assistance could work on these easy roads. Their trucks would never change lanes on their own. They would not mind going slow behind other trucks. They would not mind anything. They felt they picked the easiest driving problem to solve, other than, of course the issue of moving an 40 ton vehicle at high speed and the risks that entails.

Current fleet operators are not early adopters. It is indeed going to be hard to convince them to do something this radical, even with the problem in finding good drivers, or any drivers at all. Starsky’s original plan was all remote operation, sticking only to roads with good data. In 2015, I told them that finding such roads would be a challenge. It’s more possible today, and the deployment of services like Starlink DTLK will probably make it much easier soon.

If existing operators are not early adopters, Starsky reports that investors are scared of having to become a freight carrier in order to be a robotic truck company. They’re probably right — VC investing habits are strange to the rest of the world. They expect to invest in a dozen interesting startups each with the potential of becoming a “home run” and expecting the other 11 to try like crazy for that but flame out. Traditional businesses don’t usually fit that mold.

So here we have a case where Seltz-Axmacher is right that trucking can be a hard sell — though it hasn’t stopped many other trucking options from getting large investments.

Is the AI too hard?

You’ll need to read the essay for the full details, but it touches on one of the big questions of 2020 — how much harder is self-driving AI than people originally thought? In particular, there was a lot of enthusiasm, some of it definitely hype, about the powers of deep neural networks.

Pretty much every car team has done extensive work building training data for these machine learning approaches. That means gathering real world data (images and LIDAR clouds) and having humans label them to train the AI. This technique has delivered astonishing results, by the standards of just a few years ago. The question is, is it good enough to provide the quality level for self driving, and when will it be, and will it ever be?

Seltz-Axmacher correctly points out it’s pretty easy to get some early impressive results, and this guiles people into thinking full success is just around the corner. Several companies have even tried to build self-driving systems with “full end to end” neural networks, which are black boxes you stick camera pixels in and get driving commands out of (steering and pedals.)

He’s right that pure supervised machine learning is not enough right now, and may be some distance in the future. Tesla TSLA is betting it isn’t, but most companies are trying to build hybrids that use other algorithms combined with asking machine learning to do what it’s best at. They still believe this strategy will succeed. Generally, they have looked with disdain on those hoping to use an end-to-end approach, for the very reasons that Seltz-Axmacher outlines. In 2019 and 2020, there has been a pull-back by several players, particularly those in no particular hurry to see the automotive industry disrupted. Those who exist to do that disruption always expected the problem to be hard, I believe, but do not think their efforts are wasted.

That includes the self-driving truck world. Many companies have been attracted to trucking because highway driving is simple, even if trucks are fast and heavy. I mean really simple compared to urban streets. And the commercial value is also very clear. If anything, the commercial value is too clear, and there could be backlash when accidents come (even if at a lower rate) that people are being hurt just to make shipping more efficient, rather than to change how transportation works in general.

VCs don’t invest in this style of business

This claim is also true, though not entirely. There are forward thinking VCs and strategic investors who could be sold on a somewhat more capital and infrastructure intensive business. It’s true that, given the choice, they would rather invest in an Uber that writes only software and owns no cars. The returns are far greater. But even Uber can get investment selling a story of switching to owning large fleets of robotaxis, replacing their drivers.

It may simply be that as the slowdown and market jitters have come, it was Starsky’s plan and company that didn’t pass muster, and not the concept behind it. Certainly several other companies have raised rounds and gotten good valuations, though perhaps not as stratospheric as the valuations from a couple of years ago. There is still a very big prize to be won.

Safety seventh

Seltz-Axmacher touches on a real issue when he wonders how much people really care about safety. After all, every company in every presentation you see says, “we are all about safety” and “safety is priority #1 for us.” Now, you have to say that, and everybody is very interested in safety, because if you can’t attain safety, you can’t put your product on the market. So in that sense it’s a top priority. But in reality, in almost every business, Safety is definitely third after functionality and price. We can recount a hundred stories of products that could have been safer if they cost more money, or has less functionality. After all, self-driving cars that only went 10mph would be pretty easy to make safe quickly, but nobody would want them.

In fact, when car buyers are asked what factors they are considering in choosing their next car, they always list “safety” as the first choice. Studies of what factors actually govern their choices have suggested it’s really in more like seventh place. Otherwise, nobody would buy from anybody but the highly-safety focused brands like Volvo and Mercedes, which at various times in history have had the top reputations in that area. They don’t.

But Seltz-Axmacher points out something stronger — that the public, press and investors don’t get excited about safety because it is inherently boring. And it is. The ideal demo ride in a self-driving car is dull as dishwater. It’s hard to demo safety.

In the early days of the field, when advising a potential X Prize on self-driving to follow on the heels of the DARPA challenges, I suggested a man vs. machine safety contest. Vehicles would drive a tricky course, and fake obstacles, inflatable pedestrians and cars on small robot platforms, would create problems. Both the skilled race drivers and the robocars would compete on who could avoid hitting anything. It might have been popular — not when perfect, but when things are hit — but once the robots got perfect while the famous race car drivers were not, it would actually install confidence in the public. But nothing like this has ever been done, and no demo like this has been set up, both because nobody wants video of cars hitting even balloons, and it turns out that just driving was complex enough that handling fake situations never got high on the priority list. Teams now do this in simulator, or sometimes on test tracks, but it’s never the exciting demo. (Waymo shows a video of their car reacting to employees letting moving boxes fall onto the road.)

Is everybody doomed?

I wasn’t in the VC meetings that turned down more funding for Starsky. Today’s VC climate has cooled, and a lot of companies are being turned down. They may have had other flaws which they don’t want to go into. I suspect a lot of companies will continue to get funding, though some will be hurt by having initially received high valuations that can’t be sustained.

And it may be true that building a robocar or robotruck just isn’t a game for a small startup. It’s hard enough for the megafunded startups like Zoox, Cruise and Aurora. There’s a tremendous amount of hard slogging detail work to get from 99% to 99.9999%, which is where you need to be. It’s not 1% harder, it’s 10,000 times harder, and not everybody realizes that. The closer you get to great safety, the harder and harder it is, because each issue becomes harder to find, and each change could cause a regression on something fixed long ago. This may remain something for the big boys, at least for a few years. (Things which took billions to do the first time eventually become doable in a dorm room, it often seems.)

Some companies were going to fail. There was no way they could all survive. Indeed, there is no way that most of the teams out there will survive. That’s to be expected in something as audacious as this. Big valuations demand big results, and only a few will deliver them.


Follow me on Twitter or LinkedInCheck out my website