Tesla Autopilot launched to critical acclaim, but the hype has been tarnished by the death of an owner using the system in the USA. As an investigation into the accident that killed Joshua Brown continues, Tesla has given the driver assist system a significant upgrade. Thanks to more advanced signal processing, information from the inbuilt radar sensors can now be used to form a more accurate picture of a car's surroundings.
At the moment, radar plays a supplementary role to the camera systems in Autopilot. Although they're excellent at cutting through foggy, dusty, rainy or snowy conditions, radar sensors have a few factors limiting their effectiveness. Objects made of painted wood or plastic appear almost translucent to radar sensors, as do humans.
Meanwhile, curved metal surfaces amplify the radar signal, which means an aluminum can lying on the road could be misconstrued for a big, threatening object. Slamming the brakes on to avoid a semitrailer is a safety feature, slamming them on every time an errant can rolls across the road is annoying and dangerous.
To make sure this isn't an issue with the updated Autopilot system, Tesla's latest software can recognize up to six times more objects without any hardware tweaks. There's also a lot more information about each object within the software. Having taken a set of radar snapshots, the new software is able to assess them with the car's projected course and speed in mind. This helps it to work out whether a threat is real or just a false alarm.
Even with these steps in place, Tesla is going to continue to develop its fleet learning capabilities as the new radar heavy Autopilot rolls out. Overhead signs positioned at the top of a hill or placed at the entrance of a tunnel where the road dips away look like threats to radar.
Rather than slamming on the brakes, cars will simply note the position of bridges and tunnels without taking any action. The computer will make note of when it would have hit the brakes, along with what action the driver took. If a large number of drivers just cruise past a certain object radar thought was a threat, it will be added to a geocoded whitelist.
Should radar detect a threat, and the system data suggests the odds of a "false braking event" are slim, it will slowly increase brake pressure as it becomes more certain of the situation.
Along with this major improvement, which Tesla says would detect a UFO landing on a foggy highway, there are a heap of small tweaks to make Autopilot smoother in regular conditions - the camera can now detect indicators from cars in front as they're merging, and the car will automatically adjust its lane position to avoid drifting drivers. It'll also apply more brake pressure if the driver is being light on the pedal in an emergency, and the car can exit US highways if prompted by the indicator.
Source: Tesla Blog