Crashes are rare events.
Yes, there are nearly six million (6M) crashes every year in the United States alone. But in a non-coronavirus year the U.S. drives 3.6 trillion miles, which averages out to approximately one accident in every 600,000 miles. Given that the average American drives 31.5 miles per day, that would equate to a staggering fifty-two (52) years on average between accidents if you’re not a teenager. Normally, such an incredible statistic would be reason for joy.
However, here’s the rub: autonomous vehicles (AV’s) will need to understand how best to react in the event of a near-accident or the unavoidable accident. A deer darts from behind foliage. A high-speed vehicle emerges from an alley. A rollover accident unfolds directly ahead. These are extremely uncommon events and, therein, the best and worst driver responses to these situations are unavailable. “There’s a huge scarcity in crash-scenario data for self-driving cars,” states Peter Haas, the Associate Director of The Humanity-Centered Robotics Initiative at Brown University. “Most cars have been driven in very safe conditions with nearly-zero crashes.”
In response to this, Director Haas and Brown University have started a data collection for the masses. “We will be using very high-resolution simulation tests with human participants in virtual crashes and recording their reactions to help create a database for self-driving cars to respond to near-crash scenarios. The recorded information from the drivers would eventually help train the cars how to drive in those scenarios in the same way companies have trained vehicles to drive normally for hundreds of thousands of miles.” The scenarios are based upon the National Transportation Safety Board (NTSB) accident reports, and the simulator records how the drivers react – both good outcomes and undesirable endings –in the form of steering, braking and acceleration data in combination with the camera and sensor information.
For those techies reading this article, two other data experiments might come to mind: 1) the M.I.T. Moral Machine Experiment and 2) the “This Cat Does Not Exist” website. The Moral Machine experiment has collected millions of purely-ethical data points about how an autonomous vehicle should react in a crash. Should the quasi-robot spare the lives of the many, older pedestrians or the few, younger ones? Should the car T-bone a trolley or side-swipe a biker? But unfortunately this data is theoretical. As Haas summarized it, “This was purely people’s intention of how they’d want the vehicle to react, but it was missing the ‘impending danger’ to the driver and the potential mitigation or avoidance of the near-crashes.” The latter experiment – the “This Cat Does Not Exist” website – shows how Artificial Intelligence (AI) engines struggle with unusual scenarios (*bad pun warning: for a cat this would be a “long-tail” scenario). The website creates an amalgamated feline-like creature based upon thousands of pictures of cats, but the rare events (e.g. person cradling cat, kitten dressed in costume) create problems for the AI engine. Very typically the resulting picture has five legs or some wild defect showing that the formation of a theoretical picture using rare events can create an abomination. So the techie imagining the output of those other experiments and the underlying goal of functionally-safe, autonomous cars realizes the issue: no realistic crash data that includes the impending danger and associated successful, driver reactions.
Haas sees Brown’s intended pool of information as being the needed enabler. “Our plan is to release the database as a practical tool for companies to use, and then gauge the interest in moving forward with additional scenarios. We don’t want to get into competing on algorithms. We’re hoping this can be a resource that any company can use to test out their crash scenarios and train their system.” Presently the data collection is on hold since COVID-19 has made running human trials difficult, but the team expects to resume early in 2021 by running more subjects, expanding the scenarios and even possibly gamifying the solution to get greater data than a singular lab can generate.
When asked what inspired this effort, he smirked and stated, “My first experience with a major robot beyond middle school was at the TED conference in 2011 with the Google self-driving car. I was convinced it would be out in two years and I was incredibly excited.” His smile flattened slightly. “But at the end of the ride, I found out the sensors cost $75,000 and I realized it wasn’t coming anytime soon. Years passed and the hardware decreased in cost to under $100 with nearly the same performance. At the same time, the GPU Revolution has theoretically allowed for the computation to be quicker, lighter and smaller, but ultimately it’s the decision-making that holds back autonomy. Personally, I feel by collecting this data we’re contributing to what will eventually be a production solution that will change the world.”
Crashes are rare events.
And hopefully they’ll be even rarer in the future.