Interest in self-driving vehicle technologies has grown during the pandemic, particularly as surges in infections make deliveries and taxi rides riskier. Markets and Markets optimistically predicts that autonomous car sales will reach 62.4 million units by 2030, growing at a compound annual growth rate of 13.3%. Meanwhile, Statista expects that the worldwide driverless car market will be worth some $400 billion by 2025 — assuming that current trends hold.
But driverless vehicles remain risky propositions. As self-driving car companies like Waymo and Cruise accelerate testing of their technologies, accidents are increasing, too. Trust among consumers is unsurprisingly low, with only 17% of U.S. adults saying in a Morning Consult poll that they believe autonomous cars are as safe as cars driven by humans.
A relatively new startup, Annotell, claims to offer a solution in a platform that ostensibly “[makes] safe perception of self-driving cars possible.” By combining software with expertise to shorten the production timeline of driverless cars, the company — which says it’s secured contracts with “leading OEMs and self-driving companies” — has raised $24 million in venture capital to fuel its initiatives.
Gothenburg, Sweden-based Annotell was founded in 2018 by Daniel Langkilde and Oscar Petersson. Langkilde was previously a business analyst at Boston consulting firm Arthur D. Little. Langkilde, a data scientist by trade, spend several years at cybersecurity company Recorded Future heading up data collection and analysis projects.
“The motivation [behind Annotell] was to make it easier to use machine learning for safety-critical applications,” Langkilde told VentureBeat via email. “Annotell is designed to answer questions like, ‘How can we make sure the model we trained is good enough to be safe?’ [and] ‘How likely is it that my car can see a pedestrian at 100 meters distance if it’s raining?’”
Annotell offers a range of software to assist in developing self-driving vehicles, including perception analytics tooling that tracks certain performance indicators to measure dataset quality. The platform can connect vehicle sensors to ground truth data to validate safety requirements, enabling customers to view, sort, and manage their data in the cloud.
As they’re being developed, self-driving systems rely on data that’s been labeled to indicate objects of interest (e.g., streetlights) in images or videos, which help the system to recognize the objects. It’s an error-prone process — mistakes can crop up both during the data collection and annotation phases. A system developed only against a tropical environment will perform poorly when exposed to snow. So will a system that’s been fed data with unspecific or incorrect labels.
“Our customers typically have petabytes of data, and training machine learning models is a large part of our engineering effort,” Langkilde explained. “For us, the biggest challenge is estimating the certainty of our ground-truth being correct. Most people falsely assume ground-truth is ‘the truth.’ In practice, there is always a level of certainty that you need to estimate. We make use of a combination of sophisticated statistical models, sampling techniques, and human-machine collaboration to establish how reliable a dataset is.”
Langkilde personally believes that safety is the most important constraint in commercializing autonomous vehicles. If autonomous vehicles cause more fatal accidents, he says, it won’t just be a tragedy — it’ll risk deteriorating people’s trust in the technology.
“Cars are very different from, say, advertisements on a website. Mistakes have real consequences, and we think the software community needs to work closely with the vehicle safety community to find the right path forward,” Langkilde told VentureBeat.
But even if Annotell’s solution is as comprehensive as Langkilde claims, self-driving cars — as well as advanced driver assistance systems like Tesla’s Autopilot — face hurdles. A sobering 2020 study from the American Automobile Association found that most semi-autonomous systems on the market including those from Kia and BMW ran into problems an average of every eight miles. The “fully autonomous” system developed by Alphabet’s Waymo, one of the best-funded self-driving efforts to date, still struggles to handle certain left turns.
“The term ‘self-driving’ has lost its original intended meaning because the driving assistance feature on the cars that have been labeled ‘full self-driving’ cannot maneuver without constant human supervision,” Steven E. Shladover, a research engineer at UC Berkeley specializing in vehicle automation, wrote in an opinion piece in Scientific American. “At the pace [we] are now going, it will require decades to expand to anything approaching nationwide deployment.”
Other studies have questioned whether autonomous vehicles will be ever be the “game changer” for safety that some predict — particularly if they drive too much like people. Langkilde (who isn’t entirely neutral in the debate, to state the obvious) says it’s a tooling problem. It’s his assertion that dataset annotation platforms currently on the market like Dataloop, Tasq.ai, Labelbox, and Quality Match aren’t tailored for autonomous driving challenges.
“We’ve always believed developing safe autonomous vehicles would be extremely demanding. Our assumption since day one has been that our customers would need specialized tooling beyond what a single company can develop internally,” he said. “I think, as with all industries, autonomous vehicle companies need to be clear about what is their core business and what they should source from vendors. We are investing everything we’ve got into building the best perception performance evaluation solution on the planet. That way, we can help our customers make progress faster.”
Metaplanet and NordicNinja led the latest round in 70-employee Annotell, bringing its total raised over $31 million. Existing investors Ernström & Co and Sessan AB also participated.