Tesla calls its automated driving system Autopilot. Introduced in 2015, the system included such advanced features as allowing drivers to change lanes by tapping a turn signal and parallel parking on command. To make the system work, Tesla initially installed sonar sensors, radar and a single front-facing camera at the top of the windshield. A subsequent version, introduced in 2016, included eight cameras all around the car to collect more data and offer more capabilities.
Musk’s future vision is eventually to offer a “Full Self-Driving” mode that would replace a human driver. Tesla began rolling out an experimental version of that mode in October 2020. Although it requires drivers to keep their hands on the wheel, it currently offers such features as the ability to slow a car down automatically when it approaches stop signs or traffic lights.
In February, Tesla recalled more than 362,000 U.S. vehicles to update their Full Self-Driving software after the National Highway Traffic Safety Administration said it could allow vehicles to exceed speed limits and potentially cause crashes at intersections.
As with many artificial-intelligence projects, to develop Autopilot, Tesla hired data labelers to identify objects in images and videos to teach the system how to respond when the vehicle was on the road or parked.
Tesla initially outsourced data labeling to a San Francisco-based non-profit then known as Samasource, people familiar with the matter told Reuters. The organization had an office in Nairobi, Kenya, and specialized in offering training and employment opportunities to disadvantaged women and youth.
In 2016, Samasource was providing about 400 workers there for Tesla, up from about an initial 20, according to a person familiar with the matter.
By 2019, however, Tesla was no longer satisfied with the work of Samasource’s data labelers. At an event called Tesla AI Day in 2021, Andrej Karpathy, then senior director of AI at Tesla, said: “Unfortunately, we found very quickly that working with a third party to get data sets for something this critical was just not going to cut it … Honestly the quality was not amazing.”
A former Tesla employee said of the Samasource labelers: “They would highlight fire hydrants as pedestrians … They would miss objects all the time. Their skill level to draw boxes was very low.”
Samasource, now called Sama, declined to comment on its work for Tesla.
Tesla decided to bring data labeling in-house. “Over time, we’ve grown to more than a 1,000-person data labeling (organization) that is full of professional labelers who are working very closely with the engineers,” Karpathy said in his August 2021 presentation.