Você está na página 1de 4

by TAYLOR CONE ENG + TECH

Teaching
Credit: sxc.hu

Helicopters
to Fly

A radio-controlled helicopter,
glistening bright red in a baby
particular, they study the controls necessary
to maintain stable flight as well as perform
Researchers
blue sky, carries out stunning stunning maneuvers. With several onboard develop learning
maneuvers – rolls, flips, inverted sensors, two ground-based cameras and a
hovers, and more. The air show- few computers doing the brute force data algorithms
analysis, the control system has learned to
for helicopter
caliber routine continues for several minutes,
each maneuver executed perfectly and execute some impressive aerobatics.
without mistakes. After the show, you walk
over to the pilot to congratulate him or her Can Computers Learn aerobatics
on an outstanding performance only to find The advent of artificial intelligence brought
that a computer was in complete control for a number of revolutionary concepts to our
the duration of the show. understanding of what computers can do.
One such concept is machine learning,
Sound like a dream? Thanks to recent which involves the design and development
developments in machine learning and of algorithms that allow computers to learn
artificial intelligence this scenario is no from the patterns found in recollected data.
longer science fiction. Stanford researchers This grants computers abilities never before
have developed new algorithms that allow utilized and has implications on the areas of
computers to “learn” Through complex logic robotics, manufacturing, warfare and search
and countless lines of code, some computer and rescue among others.
systems have the ability to learn from the
patterns they observe. Conventionally, programmers write codes
that explicitly outline every command to
Under the direction of Computer Science a computer. This is useful and efficient in
Professor Andrew Ng, Ph.D. student most contexts and has had great success
Adam Coates has spent the last five years throughout the history of computing.
developing an “apprenticeship learning” However, with more complex control
system that involves a human “teacher” problems, such explicit coding is extremely
from which the computer observes and inefficient. In these situations, it may make
learns. Coates, along with fellow Ph.D. more sense to turn to machine learning
students Pieter Abbeel, Timothy Hunter, and techniques, where the computer has the
Morgan Quigley, use radio-controlled (RC) ability to determine the commands by its
helicopters as the test bed for the project. In own.

volume VIII 29
Credit: Andrew Ng

Snapshots of the flight of the For example, imagine a scenario as simple as fly the helicopter the way a human could
helicopter during a complicated picking up a pen. Think of all the commands to demonstrate that rather than solving this
maneuver. your brain sends your muscles during this using classical control methods you could
action. Your brain considers how far away use artificial intelligence algorithms to learn
the pen is, when to open your fingers, to fly this the way a human does.”
which fingers to use, how tightly to grasp
it, which muscles to employ, how much to “As soon as we had a human to teach the
contract these muscles, the speed at which computer so the computer knew where
to execute all of this, and much more. If we to start, that completely changed the ball
were conscious of every command going game,” Coates explains. Not only do they
through our brain at a given moment, we now have a human-controlled system for
would not be able to focus on anything else. the computer to observe, but the responses
– the inputs and outputs – are all from
Similarly, if we are able to minimize the the real world rather than a simulation.
amount of explicit coding we put into a Combined, these benefits make for a much
program, it will be that much easier to more effective machine learning system.
understand and control. This is exactly what
Coates and his colleagues have in mind: Why Helicopter
if they can simply write robust learning
algorithms, their programs will be able to
Aerobatics?
Although autonomous acrobatic helicopters
figure out what to do in specific situations
are the most tangible result of this project,
on their own.
it is important to note that in the eyes of
the researchers, helicopters are merely test
Apprenticeship beds for the development of their machine
Learning learning algorithms. So if they are just a test
Ng’s research group chose to pioneer a bed, what motivated the research team to
new type of machine learning for their use helicopters?
autonomous helicopter project. Rather than
using conventional machine learning that Mostly, helicopters present a unique design
involves the computer running a simulation challenge for they are inherently difficulty to
and learning from the resulting patterns, control.
they integrated a human element.
“If you’re flying in an airplane and you let
Expert RC helicopter pilot Garett Oku go of the controls...the airplane will tend
acts as the team’s teacher – not teaching to keep itself upright without any human
the students how to fly, but teaching interaction,” says Coates. The same is
The breadth of the the helicopters. Having a human teacher, certainly not true for helicopters, which tend
particularly such an experienced pilot like to flip over and crash without any human
applications of these Oku, “made us realize that when you have input. According to Coates, “they’re much
this slight advantage of having someone
learning algorithms who can help you with the learning process,
more complicated and the behavior of a
helicopter is just really difficult to predict.”
is only limited by that is a huge boon,” Coates says.
If stable helicopter flight is already
our willingness to Since humans can already fly RC helicopters
with relative ease, it seems logical to replace
difficult enough, why make it even more
complicated by pursuing aerobatics? “The
teach our skills to computer-run simulations with human pilot reason we wanted to do aerobatics is
mentors. As Coates describes, “we set out because we can demonstrate these learning
computers. to use learning algorithms to try to learn to algorithms are able to solve really hard

30 www.stanfordscientific.org
control problems that until now have been To add position information, the team has
completely inaccessible,” Coates explains. a pair of cameras on the ground as well.
“Now we can actually meet the level of a These two cameras effectively act as eyes,
human pilot, which is brand new.” tracking the helicopter in the sky and
generating the data necessary for specific
This is the most intriguing part of the position knowledge. “All of these sensors
project: humans have already solved get a coherent estimate in the computer of
this extremely difficult control problem. where the helicopter is and which way it’s
Knowing little or nothing about pointing,” Coates explains.
aerodynamics, a human pilot can fly
aerobatics with a radio-controlled helicopter. The Software
Humans “clearly have this system that’s able Considering how many sensors are
to fly helicopters well beyond anything that providing the computers with constant
the best computers were able to do, yet as input, it is a veritable challenge interpreting
a control problem it’s still largely unsolved,” it all and producing meaningful outputs.
Coates adds. What does the helicopter’s “learning
process” actually consist of?
The Hardware
Even though the project’s focus is software The first step of the process involves a
development, it is nonetheless imperative to human pilot – Oku in this case – conducting
have a robust physical platform with which several aerobatic flights with the helicopter
to test the algorithms. In fact, the physical while the computer “observes.” During
implementation can sometimes cause this time, the computer examines the
the most problems, as Coates describes: pilot’s every input and the exact result.
“We were worried that even if the learning These results vary by trial due to weather
software were good enough to do [what conditions as well as changes in the system,
we wanted], we were concerned that sensor and the computer records these minor
technology would just not be capable.” differences among the trials. After enough
Fortunately, the hardware proved to be very trials, the computer can then differentiate
successful. the pilot’s intended trajectories and the
actual resultant trajectories, thereby
As for the airframe itself, it is simply an understanding the necessary commands
off-the-shelf helicopter, as are the engines for intended results in a wide variety of
and the rest of the propulsion system. The scenarios.
rest of the hardware – the gyroscopes,
accelerometers and compasses – are all “The basic approach is to have a human fly
custom added and stored in a 100-gram the helicopter while the computer observes
black box on the helicopter. how the helicopter behaves,” Coates says.
“From that, the computer tries to learn what
Inside the black box are three gyroscopes the human was trying to accomplish, the
which sense angular rotation about three maneuvers they were trying to fly. Then,
axes, three accelerometers which sense once we put the computer in control of the
acceleration along the three linear axes, helicopter, it has to figure out what actions
and a three-dimensional magnetic compass to execute in order to repeat what the
which senses the earth’s magnetic field. human did.”
Together, these sensors give a relatively
good idea of the orientation of the aircraft How well the computer learns and
with respect to the earth. accurately replicates the maneuvers carried

volume VIII 31
Credit: Andrew Ng

The helicopter learns how to recover out by the human pilot is actually quite see the path to get there. “We could have a
its original position after flying upside- astounding. “We can fly air shows that are helicopter fly over, build a three-dimensional
down. as good as those flown by our expert pilot,” map of the area, and send it back to the
Coates claims. Pretty impressive, considering robot on the ground,” Coates describes. The
the inherent complexity of the system being robot on the ground could then figure out
used. “These maneuvers represent by far the how to get to its goal from that data.
most advanced, most difficult maneuvers
flown on any computer controlled The breadth of the applications of these
helicopter,” Ng adds. learning algorithms is only limited by our
willingness to teach our skills to computers.
Firefighting and This research demonstrates that computers
Autonomous can learn to carry out extremely difficult
Cartography “thought” processes, so it is possible that
nearly computers will eventually be able to
If the researchers pursued the possibility
mimic anything humans can accomplish.
of extending the same principles into
full-scale helicopters, these algorithms
would have great implications in tasks such The Future of Computer
as firefighting. In fact, a more complex Learning
algorithm would be necessary for such a While this project has reached (and, in fact,
scenario, which would require consideration surpassed) its goals, the field of computer
of the air currents associated with fire. learning is still young and there is much to
be done. It is not difficult to imagine several
“The interesting thing about the firefighting scenarios where machine learning would
application is you have a lot of nasty air be beneficial. Robotics, ground and air
currents moving around that can affect the vehicle control, and manufacturing, among
helicopter,” Coates explains. “Excellent pilots others, are suitable areas for applying
can compensate for this but it would be a learning algorithms that would be greatly
little scary to have a hand-tuned control transformed by such concepts.
system flying because you have no idea
what will happen when it gets hit with However, one important question remains:
turbulence.” how far will we go in granting computers
the ability to think, learn and perhaps
If the team’s algorithm were able to watch eventually feel? At this point, since we are
an expert human pilot fly in such conditions, far from mimicking human abilities, such
“These maneuvers it would learn how to react to unforeseen considerations are not paramount, and
circumstances and hopefully be able to do research continues unabated. In the future,
represent by far the the same when placed in that environment. as these algorithms become more complex
and gain broader abilities, we may need to
most advanced, most “Lately we’ve seen a lot of firefighting in
California with helicopters and we’ve seen be more cautious with our research. As for

difficult maneuvers how dangerous that can be, and it’d be


nice to get those people out of harm’s way,”
now, we can commend the magnificence of
the work of Ng’s team and eagerly wait for
flown on any Coates says. what is next.

computer controlled Another useful application is in autonomous


To Learn More
flight and aerial mapping and cooperation
helicopter.” - Andrew between air and ground vehicles. Imagine For more information, visit the website of
Credit: sxc.hu

Stanford Autonomous Helicopter at


Ng there is a ground vehicle that is trying to
reach a goal somewhere else but it cannot
heli.stanford.edu

32 www.stanfordscientific.org

Você também pode gostar