Researchers at Stanford have developed an artificial intelligence system for a helicopter, capable of learning to perform high-level aerial stunts by watching other flying helicopters
In movies along the lines of 'Terminator', artificial intelligence is described that is able to imitate the actions and actions of humans at such a high level that it is able to surpass them in various fields. There is no doubt that the films present the threat of the resurrected golem to its creator in a convincing way, and many futurists are convinced that within a few years the professional fighter pilots will be replaced by fully computerized systems. These systems will have faster response capabilities and will be able to adapt to the enemy pilot. But the sad truth is that to this day there are still no artificial intelligence systems capable of learning and developing effectively in terms of flying in the air. except one.
Computer engineers at Stanford University have developed an artificial intelligence system that allows robotic helicopters to teach themselves to perform dangerous stunts by watching other helicopters perform the same stunts. The final product is an independent helicopter capable of performing a dizzying array of complicated stunts on its own.
Writing software intended for robotic helicopters is not an easy task, partly because the aircraft, unlike the plane, is not stable by itself. "The helicopter does not want to fly. He always just wants to flip over and crash," says Garrett Oko, who specializes in flying remote control robotic helicopters. For the scientists, the helicopter is an 'unstable system' that destabilizes when it does not receive constant input. Pieter Havil, one of the graduate students involved in the research, compares flying the helicopter to balancing a long pole on the palm of your hand. "If you don't provide constant feedback, it will crash."
If so, why not design the autonomous helicopter so that it mimicked the precise finger movements of an expert pilot on the remote control? It turns out that this approach is doomed to failure due to the uncontrollable variables, such as the gust of wind. Because of this, the researchers took a different approach. Oko and other pilots put on a whole display of aerial exercises, and every movement of the helicopter was recorded. The pilots repeated each exercise several times, and the helicopter's trajectory changed slightly each time due to the different conditions. The algorithms created by the group of researchers were able to distinguish the ideal route the pilot was trying to reach, and transmitted the information to the robotic helicopter. When the robotic helicopter took off, it could perform the stunt better and more reliably than Oko himself.
There is great interest in independent helicopters, which can be used for many purposes. They can be used to search for mines in combat zones, or to map fires and locate survivors in large areas in real time. Firefighters often have to rely on information that arrived a few hours ago, but a team of independent helicopters patrolling the burning areas can solve the problem.
"In order for us to be able to rely on helicopters for important missions of this type, it is important that we have very flexible and very reliable means of controlling the helicopters, which can fly perhaps at the level of efficiency that the best pilots in the world can," says Andrew Ng, the professor who directed the project. He adds that Stanford's autonomous helicopters have taken a big step in that direction.
7 תגובות
Great, now you can attach a missile to any of the smart units and kill a few more people. It's a shame that every development has one real, military application.
Yehuda:
Your description is not accurate.
The researchers did not study the route.
Software that followed several attempts by a pilot to perform the exercise extracted the "requested" route from the flight data and transferred it to the unmanned helicopter in the circle.
It's just learning by computer. The fact that there are human-made algorithms here was clear in advance.
What I read in the article does not completely agree with the explanations I read in the comments.
In fact, even before the start of the current project, the helicopter was equipped with software that allows it - given a requested route - to take into account all the current input data on its position in the air, engine feedback and wind directions, and follow the route accurately.
The detail that seems to have been completed in the current project is the creation of the way in which the request describing the route it must take will be fed to the helicopter.
I must point out that I have seen solutions based on a similar idea already decades ago in robots built for painting cars.
Even there, it was difficult to "explain" to the robot how it should spray the paint at any given moment - from where, in which direction and for how long - so it was taught to simply repeat the movements of a person who "trained" it (that is, the person held the paint sprayer and painted the the car and the robot only learned the position and direction of the sprayer at any moment). After the learning phase, the robot could perform the work by itself thousands of times while saving manpower.
Here, of course, the general problem is more difficult because the plane has to take into account the changing wind data, but according to what I understand from the article - this difficulty was actually solved before and made the current stage much simpler and very similar to what was done in the past.
The title of the original article is slightly different and it is
Swarms of robotic helicopters developed at NASA developed superintelligence on their own and managed to transfer it without human intervention to seven pilotless planes that flew themselves to an unknown destination
higgs,
Thanks for the completion.
Too bad the point is missing……..
This is what happens when you try to penetrate the world of the professionals, for some reason the main issue is left out.
I will allow myself to complete.
The generalization here reveals that non-linear systems in their nature are able to take advantage of this very feature to gain an advantage in solutions for performing complex courses of action. Those under normal conditions are considered difficult to solve in the first place due to the basic non-linearity.
In fact, the algorithm somewhat surprises its authors in that it is actually much simpler in its structure than expected.
There are many new mathematical fields that take advantage of properties of this kind, for example RMT.
"Pretentious" titles are meant to arouse the curiosity of the person in us.
Brain replacements in order to free us from too "heavy" tasks.
But, the responsibility... to think... and clarify, I don't think anyone would really want it taken away from them.
.
Sahatan.. on the corner of the new knowledge ran to our left..
The title of the article is a bit pretentious:-"
"A robotic helicopter can learn to fly by watching its friends"
This is not exactly what we understood from reading the article:-
"The algorithms created by the group of researchers were able to distinguish the ideal route that the pilot was trying to reach, and transmitted the information to the robotic helicopter"
In other words, researchers studied the movements of other helicopters and fed the information into the new robotic helicopter that turned out to be more successful. He did not learn from her, rather, they taught him. And that's a big difference.
This does not detract from the importance of the researchers' work, but a little precision in the publication.
Good Day
Sabdarmish Yehuda