demonstrations are often suboptimal in different ways,” their Learning for Control paper noted, “suggesting that a large number of suboptimal expert demonstrations could implicitly encode the ideal trajectory the suboptimal expert is trying to demonstrate.”
They discovered that merely using an arithmetic average of the states observed at any given time in the expert demonstrations would fall short of arriving at the desired trajectory, explaining that, in practice each demonstration would occur at different rates, and hence make impossible an attempt to combine states from the same time-step in each demonstration.
However, by employing the machine learning algorithm—which includes an extended Kalman filter and a dynamic programming algorithm— the researchers were able to infer the intended target trajectory and time alignment of all the demonstrations. And, while real-time variables such as the state of the air around the craft, rotor speed, actuator delays, and the behavior of the helicopter’s onboard avionics contribute to an extremely complex environment that cannot be modeled accurately, these variables can be mitigated if the programming is able to make the helicopter fly the same trajectory each time. If so, the errors caused by these variables will tend to be the same, and therefore can be predicted more accurately.
In addition to the aerobatic results of the project, Coates says the ramifications for machine learning theory
go deeper. “One of the reasons people liked our paper is that it was an off-the-shelf machine learning algorithm and we solved a strange little application nobody had thought of before,” Coates notes. “People know how hard this is, and to see that AI people solved this, I think has made a big impact. We had been preaching for a while that AI is the key to solving really hard problems that aren’t accessible to us when we’re using lots of classical methods—and if you come up with a problem and make such large strides, it really adds some weight to the argument that AI can be real and practical with algorithms that solve really hard problems.”
The future of autonomous helicopters might be even more profoundly affected by the march to increasingly powerful processors and smaller form factors.
“One way to avoid safety troubles is by making the helicopters smaller, so there are a lot of efforts going into miniaturizing the machines,” says Eric Feron, professor of aerospace software engineering at Georgia Tech University, who studied autonomous helicopters while a graduate student at MIT. “That’s where I think things are going now.”
Coates says the breakthrough Stanford research was greatly facilitated by increased processor capability that allowed real-time instruction every 20th of a second, which was not possible even five years ago. Additionally,
the advent of microelectromechani-cal systems-based sensing technology, such as gyroscopes, accelerometers, and magnetometers, is leading to increased miniaturization.
Navigationally, academic researchers are now also concentrating on developing obstacle detection technology that will allow autonomous helicopters to fly safely in urban areas teeming with tall buildings, overhead wires, and light poles. Such uses are not on the near horizon, however; the ongoing safety concerns probably point to deployment in sparsely populated areas for natural resource mapping, forest firefighting, and marine search and rescue. Human-generated mapping at quarter-meter resolution can cost $20,000 per square mile, for example, while autonomous helicopters could probably deliver the same results 10 times cheaper, says Amidi.
Georgia Tech’s Feron says autonomous helicopters will continue to offer researchers an excellent platform for further research in robotics, whether the researcher is an “aeronaut” who will still be utilizing them 10 years hence, or instead testing a more universally applicable methodology on the machines, and that wider deployment will indeed follow at some point.
“The safety and reliability issues are not unworkable,” Feron says. “I think it’s just a matter of time.”
Gregory Goth is an oakville, ct-based writer who specializes in science and technology.
© 2009 acm 0001-0782/09/0600 $10.00
Computer science was well represented when the american academy of arts & sciences (aaas) recently announced the election of the 2009 class of fellows and foreign honorary members. the 212 new fellows and 19 foreign honorary members—including scholars, scientists, jurists, writers, artists, civic, corporate and philanthropic leaders—come from 28 states and 11 countries and range in age from 33 to 83. they join one of america’s most prestigious honorary societies and
a center for independent policy research.
“since 1780, the academy has served the public good by convening leading thinkers and doers from diverse perspectives to provide practical policy solutions to the pressing issues of the day,” said leslie Berlowitz, aaas chief executive officer. “I look forward to welcoming into the academy these new members to help continue that tradition.”
Elected in the category of
computer sciences (including aI
and information technologies):
• John seely Brown, Deloitte Center for Edge Innovation/ university of southern California
• mary Jane Irwin, Pennsylvania state
John seely university
Brown • maria Klawe, harvey mudd College
• ray Kurzweil, Kurzweil technologies
• michael sipser, mIt
• alfred Z. spector, Google
• Jennifer Widom, stanford university.
Elected in the category of business, corporate, and philanthropic leadership:
• John Doerr, Kleiner, Perkins, Caufield & Byers.
In an email interview, John seely Brown offered this career advice for young people: “nurture a disposition that embraces change and that encourages you to challenge your own assumptions and having others challenge yours.”
PhotograPh by J.D. Lasica
References:
Archives