4·
1 day agoEssentially what they do is have one human operator watching ~20 vehicles at the same time. When one of the vehicles gets into a situation that the automation can’t handle, it alerts the operator to take manual control. The cars aren’t being remotely operated 100% of the time.
In order for these vehicles to operate, they must all be surveillance platforms. They are covered in cameras, microphones, radar, sonar, lidar, and other types of sensors. The more of these vehicles are on the roads around you, the more real-time surveillance data is being collected. And you better believe those cameras are doing license plate recognition of every other vehicle they pass, in case there’s any incident that would involve an insurance claim. Basically they are mobile Flock on steroids.
Training machine learning models requires data. The edge cases have the least data for training, because they involve unusual weather or road conditions, specific lighting situations, weird/poorly designed intersections, or other uncommon circumstances. Edge cases will never be trained out because there will never be enough data for each specific case to have a meaningful effect on the self-driving model. Edge cases present the highest likelihood of injury to humans, whether as passengers, other drivers, or bystanders. People will continue to die, and somewhere in a meeting room someone will be presenting calculations to a board about whether it’s worth investing in better safety controls for their vehicles, or just bribing politicians to shield them from regulations and lawsuits.