Self-driving cars face a major figurative challenge: anticipating unexpected situations and responding in a level-headed way to ensure everyone stays safe. Yet they also face a rather bizarre literal obstacle in the form of an enormous Looney Tunes-inspired fake wall painted to resemble an open road, designed to fool the vehicle into a collision.
While giant fake walls are unlikely to appear in real life, that did not stop Mark Rober—an ex-NASA engineer now on YouTube—from testing how self-driving vehicles react to cartoonish scenarios.
In his most recent video, “Can You Fool a Self-Driving Car?” he compares two autonomous driving systems: Tesla’s camera vision-based autopilot and another setup that relies on Light Detection and Ranging (LiDAR). Each system is put through a series of tests, ending with the same classic tactic Wile E. Coyote used to try to stop the Road Runner in Looney Tunes.
At about 40 miles per hour, Tesla’s Autopilot plows right through the faux wall, leaving a comically large hole behind—its third failure out of six tests. Earlier experiments tackle whether a self-driving vehicle might hit a child under poor visibility, such as fog or water sprays. During those demonstrations, the Tesla autopilot successfully stopped in front of a stationary child dummy standing in the middle of the road with no visual obstructions.
However, once the dummy was covered in fog or covered by water jets from different directions, the Tesla ran right through the dummy. In contrast, the LiDAR-based autopilot on the competing car stopped every single time. That outcome may not shock viewers, since the clip quietly doubles as a promotion for LiDAR technology. Plus, LiDAR is a much more expensive and sophisticated environment-scanning technology.
Still, the LiDAR system’s success is striking, especially now that Tesla has very publicly decided to rely on cameras alone. The argument for skipping LiDAR usually cites its high cost, the extra processing it demands, and the belief that leaning on LiDAR could slow progress in purely camera-based solutions.