Monthly Archives: January 2019

Still unsafe at any speed?

waymo one

Mindful of the fact that most of the posts on this blog are provoked by media coverage that works to further mystify AI/robotics, I thought I would break with that pattern to recognize a recent story that, to my reading, breaks with the pattern. A piece by Andrew J. Hawkins in The Verge reviews the state of robot taxis in Phoenix, Arizona. Having been assured on several occasions by tech-savvy friends that self-driving cars are in everyday operation in Phoenix, this story helps to clarify the actual state of the technology. And Hawkins raises some welcome questions as well about the claimed benefits of autonomous vehicles; questions that should be at the forefront of discussion about public investment in transportation infrastructures going forward.

While it’s not until paragraph nine that we discover that the autonomous vehicles deployed in the Waymo One taxi service actually still include a human ‘safety driver,’ there’s much to learn here. The piece is headlined by a video report from Hawkins on his own experience of the service. Hawkins points out that Alphabet/Google’s Waymo is the longest running and most extensive of the autonomous vehicle projects, with the lowest number of recorded “disengagements,” or events in which the human driver has to take over the wheel. The current trial is limited to four towns in the greater Phoenix area, and to voluntary members of Waymo’s “early rider” program (with requisite non-disclosure agreements and, we might assume, liability waivers.) We might note in the aerial views of the designated areas the Arizona landscape’s flatness, and of course we know that the reason for the old prescription to “ship your sinuses to Arizona” (familiar at least to TV watchers of my generation) is that state’s relatively rain free climate. There’s not much discussion of Arizona’s particularities here or in the media more generally, but they point us towards the question of what environmental conditions are required for the self-driving car’s successful operation.

Hawkins very helpfully introduces us to the infrastructure of sensor technologies that make the autonomous vehicle possible (the car’s sensor view is live streamed for the passenger, in part presumably to make the ride less boring). As we watch we begin to get a sense of how the car/environment might be a more apt unit of analysis than the car alone. Surrounding vehicles and pedestrians are rendered as categorically color-coded, edge-detected objects. As Hawkins compares the experience to “being in the back seat with a very cautious student driver,” we get to sit through an “unprotected left turn” (what we human drivers would call turning left mid-block onto a side street, by waiting for a break in oncoming traffic). We see how the Waymo One turns only when the change in the traffic light at the intersection ahead creates a clear, clean break in the traffic. Well worth the wait, I suspect most of us would say, though a source of reported frustration for other human drivers. For Waymo this breakdown in driving tempo poses the challenge of developing its software to enable the car to drive “more organically, more like a human.” We get a sense here, for better and worse, of the difference between an operation based entirely on metrics and algorithms, and a practice based on embodied experiences of space and time. Nonetheless, Daniel Chu, Director of Product from Waymo, translates aggregated time on the road and associated statistics into a characterization of each Waymo One vehicle as equivalent to “the world’s most experienced driver.”

Perhaps the most welcome moment in Hawkin’s account comes when he turns to the question of whether the car is actually the most imaginative, or even desirable, vehicle for the future of transportation. Millennials, he reports, indicate in poll data some doubt about the future of car travel, and a preference for better public transportation, along with safer spaces in which to bike and walk. While touted as a remedy to the proven fallibility of human drivers, comparable safety statistics for the driverless car aren’t really available, according to Sean Sweat of the Urban Phoenix Project, given the relative size of data sets on cars driven by humans and driverless cars over time. Sweat points out that the question of driver safety also sidesteps the question of how the design of urban spaces, particularly streets, might contribute to pedestrian fatalities or, alternatively, to their avoidance. This points to the much larger issue of how the investment in a future of self-driving cars might drive the reconfiguration of transport infrastructures required to enable them. Not only the cars themselves, but our roadways and urban landscapes will likely become further instrumented in the service of vehicle autonomy. This isn’t inherently a bad thing; the Copenhagen subway system, for example, has evolved through a thoroughgoing makeover of the city infrastructure to create a driverless and extremely safe transport system to accompany its bicycle-friendly streetscape (most obviously, there’s no open access to the track, even from the station platform). But in car cultures the financial expense required to re-engineer highways and cities in order to make them autonomous vehicle friendly is accompanied by lost opportunity costs, beginning with a sidelining of discussion about alternative possibilities. Meanwhile the future of autonomous cars that can drive any road, under any conditions, may take decades, Hawkins concludes, or may never happen. Far from inevitable, then, the driverless car is a project urgently in need of braking, to open a space for more innovative ways of thinking about safe and sustainable transport.