By Thomas H. Davenport
Among my most interesting jobs is being a digital fellow at the MIT Initiative on the Digital Economy. There are often interesting developments happening there on the topics of artificial intelligence (AI), digitization, and digital platforms. That was certainly the case on March 8, when I attended the MIT Disruption Timeline Conference on AI and Machine Learning. There was interesting content on a variety of topics, but a primary focus was on when specific AI capabilities might become generally available. One particular technology addressed was autonomous vehicles. The key question was when 50 percent of vehicles on U.S. roads would be fully autonomous.
MIT has some impressive faculty and alumni who have been working on aspects of this topic for a long time. One alumnus and former MIT faculty member (also a former professor at Olin College, Babson’s sister school in engineering) who spoke and participated on a panel was Gill Pratt [far right in photo], now head of the Toyota Research Institute. The panel also included John Leonard [second from left] and Tomaso Poggio [second from right], MIT faculty who have worked on this issue from the perspectives of mechanical engineering and brain research, respectively. If that weren’t enough, Manuela Veloso, head of the Machine Learning department at Carnegie Mellon University (CMU) and a robotics expert, came over from Pittsburgh to add to the panel with her insights [far left in photo]. The panel was moderated by MIT director, Erik Brynjolfsson [center in photo.]
Overhyped or On the Way?
Of course, these experts are as excited as anyone about the potential for autonomous vehicles, but they are also relatively conservative about when the 50 percent standard might be reached. If you listen only to the Silicon Valley Industrial Complex, you might think they are just around the corner. But Pratt wouldn’t give a specific prediction, Leonard said that autonomous vehicles are over-hyped, and Poggio flatly predicted it would be 20 years before fully autonomous vehicles are on the road in large numbers.
Why so long? Leonard illustrated the problem with some video footage about trying to make an “unprotected left turn” into traffic on a snowy day in his Boston suburb. There’s a long line of cars that he wants to turn left into, and who’s going to let him in? And would a machine vision system be able to distinguish oncoming vehicles and lane markings in the snow? Not anytime soon, he believes.
Poggio, who is also on the board of directors of the Israeli machine vision company Mobileye (acquired the next day for more than $15 billion by Intel), focused on the problem of identifying pedestrians. He noted that Mobileye has been able to double its ability to successfully identify pedestrians every year for the past 20 years, but the technology still wasn’t good enough.
Even more alarming, however, the improvement is beginning to plateau. He and the other panelists noted that the problem with autonomous vehicles isn’t the vehicle itself, but rather, predicting the behavior of other motorists and pedestrians.
One of Veloso’s research foci is cobots, or collaborative robots. She argues that we might be better off pursuing augmentation from smart vehicles than pure autonomy. The level of ambition for AI shouldn’t be too high. For example, in the cobots her lab has developed at CMU, the researchers wanted the cobots to be able to deliver documents around the building. Properly functioning arms that could carry documents were difficult to build, so they just put baskets on the robot. If the robot needs to get into the elevator, instead of pressing a button, it simply stands in front of the elevator and asks passing humans to push the up or down button.
Pratt from Toyota was similarly conservative in his views about full autonomy, but his research center has some worthwhile alternative goals. He pointed out, for example, that although less than 1 percent of adult deaths in the U.S. are from auto accidents, 35 percent of teenage deaths are. So Toyota is trying to develop a vehicle with a “guardian” mode to protect teens (and other bad drivers, presumably) from making lethal driving mistakes. The company is also working on a “chauffeur” mode for older drivers who need continuous help — particularly important in Japan, with its rapidly aging population.
Pratt also referred to two explosive factors that make it difficult to predict how rapidly autonomous vehicles will improve (described in an influential article he wrote in the Journal of Economic Perspectives that is republished here). One is the very fast rise of deep learning technology — critical for vision and perception. Another is the prospect of robots and vehicles whose intelligence is primarily outside themselves in the cloud and communicating all the time. In cars, this would allow for “fleet learning” in which cars on the road learn from each other.
I asked these experts a question about how long it would take to develop standards by which cars would communicate with each other. They said that the good news is that a lot of car companies and their suppliers are working on such standards. The bad news is that they are all somewhat secretive and only somewhat collaborative — as usual, everybody wants to gain competitive advantage from standard setting. This is another factor that could slow things down.
My conclusion from all this is that we are still going to have a fully autonomous future, but that we should all take a breath and relax about it, because it’s going to be a while.
This belief is consistent with the experience of “early adopter” companies. Uber, for example, has supposedly autonomous vehicles running around three states. But in their first 20,000 miles of driving, the Uber drivers had to intervene an average of once per mile.
In the short run we should be thinking about ways to provide increasing levels of driving help while still maintaining the driver’s attention — avoiding what academics call the “vigilance decrement” problem. As a human, I find it somewhat comforting that our role as drivers isn’t quite obsolete. And given our records as (not so) safe drivers, we need all the help from machines we can get.
Tom Davenport, the author of several management books on analytics and big data, is the President’s Distinguished Professor of Information Technology and Management at Babson College, a Fellow of the MIT Initiative on the Digital Economy, co-founder of the International Institute for Analytics, and an independent senior adviser to Deloitte Analytics. He also is a member of the Data Informed Board of Advisers.
Watch the video of this panel here.
This blog first appeared in DataInformed on March 22 here.
Also read Tom’s latest blog in HBR on AI here.