PHOTOGRAPHY:PETERPRATOSTART
To stop a rogue
Donkey Car, Roscoe
says, “Just put your
foot in front of it”In a warehouse in Oakland, California, a small crowd watches Will Roscoe tap a
phone with his thumb. At his feet, a radio-controlled car starts to move around
a racetrack – with no further input from Roscoe. The Frankenvehicle, which has a
camera and electronics zip-tied to its top, is called a Donkey Car. Roscoe is no
AI expert, but his creation uses neural network software similar to that which
Waymo’s street-legal autonomous minivans rely on to perceive the world. The car
learns to drive by watching a human steer the vehicle during demonstration runs.
A civil engineer by training, in 2016 Roscoe ran for a seat on the board of the Bay
Area subway system. He pledged to replace trains with self-driving electric buses,
but finished third. Building his own autonomous vehicle seemed a way to show
voters that the technology wasn’t pure fancy. “I wanted
to demonstrate it can work at a small scale,” he says.
He met fellow tinkerer Adam Conway, who built the
vehicle and its autopilot using Google’s TensorFlow
software, and after putting their designs online for others,
Donkey Cars of all shapes and sizes are now racing
around in Hong Kong, Paris and Melbourne. Tom SimoniteThe wacky
self-racer
Will Roscoe has built a cute radio-
controlled buggy with a difference:
the Donkey Car can drive itself
using neural network software