The company claims wireless transmitters on cameras used by a TV crew caused interference that disrupted signals from a transponder on the traffic light.
Intel's Mobileye has experienced an embarrassing failure after a prototype vehicle ran a red light in Jerusalem during a press demonstration.
Local news outlet Channel 10 captured a video of the autonomous car approaching a red light and continuing to pass through without any intervention from the human safety driver. The car was demonstrating Mobileye's camera-only autonomous systems, which lack expensive lidar and radar systems.
Channel 10 was told that "our crew's GoPro disturbed the signals -- and that further tests are required to avoid this [sic] kind of mishaps."
In a statement to Bloomberg, Mobileye clarified that the television crew's cameras created electromagnetic interference that disrupted signals from a transponder on the traffic light. The report suggests Mobileye was relying on vehicle-to-infrastructure (V2I) technology for traffic light sensing. Mobileye says its cameras saw that the light was red, but proceeded anyway due to the signal from the transponder.
The explanation seems suspicious for a few reasons. It is plausible that a GoPro's Wi-Fi transmitter may have caused interference with an unshielded prototype receiver, but extremely unlikely the GoPro created a false signal indicative of a green traffic light.
If the traffic light only transmits a status signal when the light is red, and the GoPro effectively jammed the signal from within the car, it seems like a glaring oversight for Mobileye algorithms to interpret the lack of a clear red-light signal as an indication that the light is green.
"It was a very unique situation," said Mobileye CEO Amnon Shashua. "We'd never anticipated something like this."
GoPro cameras are not exactly specialized camera equipment, with nearly five million units shipped globally from 2014-2017.
The incident raises more doubts about the viability of V2X technology, demonstrating that Mobileye's cars could have easily performed properly using camera sensors and traffic-light recognition algorithms alone.