This post has been corrected.
Tesla’s new autopilot features became available just a few weeks ago, and the software is already teaching itself to drive better.
As Quartz noted when the Model S autopilot features came out, the car isn’t completely autonomous, and still occasionally requires hands on the steering wheel. And the automaker is using those human interventions as “teachable moments” to make the software better.
One driver wrote:
So far I have a little over 300 miles on autopilot, mostly 20 miles at a time on my commute to and from work. The first day when I was in the right lane, as I approached exit ramps, it would dive for the exit ramp. I quickly learned to apply torque to the wheel to hold the car on the interstate until I had passed the exit.
Each day the system seems to have less tendency to follow the exit ramps as I pass. The last two days it only gave a momentary wiggle and moved over maybe six inches towards the exit ramp then it recovered and moved on down the road.
This morning it gave only a very slight hesitation, so little that I did not have to correct it at all. I find it remarkable that it is improving this rapidly.
The data is automatically compiled and uploaded to Tesla’s “fleet learning network,” so that all its cars get access to the same improvements. “When one car learns something, all learn,” said Tesla CEO Elon Musk.