As the automotive industry inches slowly ahead on the road to self-driving vehicles, we’re seeing the emergence of startups aiming to fill in some of the technical gaps in autonomous systems as they exist today. In the latest development, Annotell, a startup out of Sweden that makes software to assess the performance of autonomous systems’ perception capabilities, and how to improve that, is today announcing that it has raised $24 million to expand its business.
Daniel Langkilde, the co-founder and CEO of Annotell, likens what the company does to “a vision exam for cars, for them to get their drivers license, just like you might take a test to determine if you are fit for driving,” he said in an interview. “Annotell’s platform helps you understand the system’s performance and raise it. We guide our customers on how to improve it.” That is to say, Annotell’s products encompass analytics that test and measure the quality of a company’s data, and “ground-truth” production to improve those data sets.
The aim, he added, is not perfection but predictability, just as important for the semi-autonomous platforms (e.g. advanced driver assistance systems) that exist already today as for the fully autonomous cars that many are hoping to build for the future. “The system may not always be right, but you need to know what it can or can’t do, in order to use the system safely.”
The Series A round is being co-led by Metaplanet — the Estonian VC headed by the co-founder of Skype Jaan Tallinn that most recently also invested in Starship Technologies and was an early backer of Google-acquired DeepMind — and NordicNinja — a Japanese-backed deep tech investor. Previous backers Ernström & Co and Sessan AB also participated. Gothenburg-based Annotell has now raised $31 million, and it’s not disclosing valuation, but for some context, its customers include a number of the world’s biggest carmakers, their main suppliers and the big pure-play self-driving car companies.
The gap in the market that Annotell is looking to fill is a pretty critical one: Autonomous systems are built on huge troves of driving data and machine learning used to process that information to “teach” those platforms the basics of driving.
Using computer vision, those systems in turn can recognize red lights, or a stopping car, or when to make a turn, and so on. The problem is that these systems’ responses are based on the data that they have been fed. Autonomous systems typically can’t “reason” and make the leap to decide how to respond to an unknown variable, such as those that a vehicle will inevitably encounter in the real world.
“Machine learning is bad at processing rare but important things,” Langkilde said.
Langkilde who co-founded Annotell with Oscar Petersson — both are physicists who specialize in deep learning — said he encountered that problem when he previously worked at a different company, the threat intelligence startup Recorded Future, where he was tasked with gathering intelligence data to feed and teach the platform to better identify threats. Malicious hackers are precisely focused on finding gaps to create vulnerabilities, and that effectively upended a lot of the work his team would do to identify patterns to mitigate future attacks.
“It highlighted the limitations to me of brute force machine learning when you are doing mission-critical work,” he said.
Autonomous driving systems face much of the same issue, but it’s even more critical to get right, not least because there are lives at stake if something goes wrong. This also brings in more levels of safety and control that companies need to pass through to bring their products to market, and get consumers to trust and subsequently buy and use them.
“For people to trust machine learning and AI we have to take safety very seriously,” he said. “There is a huge difference between making the wrong recommendation on a film service and running a stop sign or running into someone. We also take that seriously. That’s why we wanted to focus on the problem.” The extra layers of safety regulation, meanwhile, also point to specific use cases and market opportunities for Annotell: It’s not just about improving systems for its customers, but creating a body of data that agencies and regulators can also rely on to give a particular product the clearance to be used.
Annotell’s approach to complementing what machine learning can teach systems is as progressive as autonomous systems are today: in part it tests and formalizes the limitations of systems that by their nature are not designed to be fully autonomous (these are the systems we have today to assist, not replace, drivers). Over time, he said, fully autonomous might also incorporate other kinds of AI approaches, such as the Bayesian Networks that are used to build causal inference algorithms. (A causal AI startup we covered last week was more dramatic, claiming that causal AI was the only hope for self-driving to become a reality, although even then it’s a big leap and will take a lot of time to come to fruition.)
For now, though, the startup is focusing its tech on safety of systems with any degree of autonomy already built in, a massive opportunity.
“Ensuring safety is the main constraint when it comes to commercial deployment of autonomous vehicles, and Annotell has made great progress in a short period of time,” said Jaan Tallinn of Metaplanet, in a statement. “We’re impressed by their software as well as the team that built it and we’re thrilled to be with them on this journey.”