As Oku repeated a maneuver several times, the trajectory of the helicopter inevitably varied slightly with each flight. But the learning algorithms created by Ng's team were able to discern the ideal trajectory the pilot was seeking. Thus the autonomous helicopter learned to fly the routine better—and more consistently—than Oku himself.
This is incredibly cool work -- it's clear how this is a more general approach than hard-coding a bunch of rules about how to fly a particular helicopter, under particular assumptions about the environment and the desired maneuvers.
In the future, I wonder if it would be possible to improve the learning algorithms to the point that no human expert is even needed: you could toss the helicopter up in the air, and have it essentially figure out how to fly the machine before it hit the ground.
This work is also a nice counterpoint to those who question the value of the work being done in academic computer science. I think there is more innovation and long-term value in these sorts of projects than in a dozen typical Web 2.0 social networking startups.
This is incredibly cool work -- it's clear how this is a more general approach than hard-coding a bunch of rules about how to fly a particular helicopter, under particular assumptions about the environment and the desired maneuvers.
In the future, I wonder if it would be possible to improve the learning algorithms to the point that no human expert is even needed: you could toss the helicopter up in the air, and have it essentially figure out how to fly the machine before it hit the ground.
This work is also a nice counterpoint to those who question the value of the work being done in academic computer science. I think there is more innovation and long-term value in these sorts of projects than in a dozen typical Web 2.0 social networking startups.