Boffins produce aerobatic copycat-copter pilotware
Computer see, computer do
Researchers at Stanford University have developed technology which lets computers handling remote-control helicopters achieve complex manoeuvres by copying a human pilot. Having "seen" a move carried out successfully once, the pilot-ware can then repeat it more consistently than the human.
Stanford grad students led by Professor Andrew Ng refer to their new methods as "apprenticeship learning". The Uni press release and resulting coverage has followed a "helicopters teach themselves by watching" line. However - at present, anyway - it would be more accurate to say something like "human pilot programs computer autopilot using sensor-equipped helicopter".
The Stanford kit works using normal, small remote-control copters fitted with a lot of sensors and a data link. First, a human pilot - in this case, one Garett Oku, a remote-control chopper ace - flies a manoeuvre. The sensors continually measure the copter's position and attitude, creating a precise record of what the helicopter is actually doing.
Previous attempts at writing code which could fly difficult manoeuvres such as knife-edges, Immelmanns, inverted tail slides etc would fail, because the programmers only knew in general terms what these manoeuvres consisted of. They had no idea of exactly what values and rates of change in velocity, pitch, yaw and roll would mean success in any given feat.
Oku, however, did know - but would have had difficulty describing his knowledge in terms of numbers. So the sensor-equipped copter was used as an interface to his brain, creating records of what he was actually doing.
Once the records existed, the Stanford pilot-ware running in a ground computer had a go at duplicating Oku's efforts. Again, the sensor-equipped helicopters were used, so that the computer could tell how it was doing.
According to Stanford, the computer was soon able to "fly the routine better — and more consistently — than Oku himself".
The machine was even able to execute Oku's piece de resistance, the so-called "tic toc". Here's a YouTube vid of the hands-off copter aerobatics:
(You'll need Flash installed and your net admins will need to agree you should be watching YouTube at work.)
For the Stanford experimental setup, the computer and some of the necessary instrumentation are on the ground, not in the aircraft. But Ng and his cohorts say that larger aircraft could carry all the needful gear themselves, producing an autonomous system.
This isn't so much a case of computers learning to do tasks, then, as of using sensors to describe an aircraft's motion in terms of numbers. Given an accurate description of what's wanted - generated in this case from previous successful manoeuvres - and a continual description of what's actually happening, the computer can easily enough line up the two.
It's an interesting trick, which could make the programming of some future autonomous aircraft easier and simpler. But it might struggle to find relevance. There have been aircraft flying for decades now which humans can control only with automated help. That help is generally provided using onboard computer and sensor rigs not unlike the Stanford kit just proven.
Flight tasks, profiles and such these days are usually developed from models and simulations and wind tunnels, not by flying real airframes under direct human control. The numbers may be refined somewhat in flight testing, but in the case of aircraft intended for automated control there would be no need for humans to get hands on in order to explore the limits of the envelope. A human pilot doesn't really add anything to the performance of the Space Shuttle or the B-2 bomber. Automatic systems can already accomplish most tricky piloting feats as well as humans - landing, air-to-air refuelling, carrier deck landings and so on.
The example of small helicopter aerobatics may actually be one of the last cases in aviation where the Stanford human-recording trick is applicable, rather than one of the first. And indeed, the ability to do airshow stunts isn't all that handy in and of itself. It isn't clear how this sort of thing is going to help helicopters search for landmines or track wildfires, as the Stanford boffins suggest.
It's still a pretty cool trick, though; and things like this often turn out to be useful in ways you wouldn't expect. That, after all, is what academic research is supposed to be about - things whose use isn't immediately obvious - and there are other fields of activity aside from aviation, too.
Sponsored: Benefits from the lessons learned in HPC