Starsky Robotics Hits A Dead End

Five years ago it seemed like every other software engineer in Silicon Valley was loading up on some venture capital money in hopes of striking it rich with automated driving (AD). Right about now we were all supposed to turn over the driving task to computers. Elon Musk even promised full self-driving on all his vehicles by 2017, 2018, 2019, 2020... Needless to say, the task has proven a lot harder than most people thought and Starsky Robotics is the latest of what will likely be dozens more AD companies to turn off the lights. 

In a blog post on Medium, founder and CEO Stefan Seltz-Axmacher announced that Starsky Robotics was shutting down after 5 years of effort trying to build robust automated trucks. Around the time Seltz-Axmacher started the company, there was enormous enthusiasm building in the tech community for a variety of new software approaches to achieving artificial intelligence including deep learning and supervised machine learning. 

In the early part of the last decade, enormous progress had been made in improving the way that software could see and draw semantic understanding of the world. Unfortunately, what Seltz-Axmacher and many others have discovered is that the progress would not continue indefinitely. 

The supervised machine learning approach used by Starsky and so many others relies on huge quantities of accurately labeled and annotated data. It’s crucial for the labeling to be accurate and for the data sets to be diverse. The data is then fed into ML models and the parameters in the model are adjusted as it “learns” what all of the objects are until the output matches the input labels. 

Think of this, like teaching a toddler colors or letters or numbers or animals. You sit down with them with a picture book and point to a red ball and identify it verbally. That’s the labelling and annotation part. Repeat that process until they child can look at it and tell you what it is without prompting. 

The problem is these models still don’t really function the way the brain does and frankly we still don’t have really a firm grasp on that or how to replicate it. The ML models seem to improve up to a point and then stall. They also don’t seem to be as good at filtering signal from noise as humans are, that’s why you see machine vision systems get fooled by a sticker applied to a stop sign. 

In his blog post, Seltz-Axmacher discusses the challenges of trying to break through these ML barriers and speculates on where AD systems are today relative to human drivers. To his credit, he emphasizes the need to ensure that these systems really are safer than humans and how difficult that is to do. His approach was to focus on trying to deploy where there was a very high degree of confidence that the system could work safely and reliably. 

Unfortunately, the venture capital community didn’t seem satisfied with that and wanted much faster growth and much higher margins than are going to be possible for many years to come. Following the terrible IPO performance of Lyft and Uber last year and then meltdown of WeWork it became far more difficult to raise more money.

Starksky’s most notable competitors right now are TuSimple and Waymo and they both have a lot more money and runway. However, throughout the AD sector, there is a recognition that reaching viable automated driving on a large scale is going to take many more years, Elon Musk’s claims notwithstanding. 

Starsky Robotics isn’t the first AD company to shut down and absolutely won’t be the last. The current coronavirus pandemic and the ensuing global recession that is pretty much assured will likely see many more give up before 2020 is over.  Those that remain are largely all using a mix of different approaches including ML and more traditional rules based software algorithms. What is clear is that there is no silver bullet to automated driving.