Massachusetts Institute of Technology created a “cheetah robot” and taught it to jump

Do you know that Massachusetts Institute of Technology (Massachusetts Institute of Technology) is there a cheetah robot? So did we, but we not only found that the project was real, but the same cheetah robot learned to jump. And you thought that humanity had stopped in robotic sniper dogs, Definite?

Read it too

publicity

The novelty comes from the Massachusetts Institute of Technology itself, because humanity obviously has the right to brag. creations that may one day kill her… In an official statement from the institute, the researchers say that “four-legged robots have been inspired by the movements of cheetahs and other animals, making tremendous progress, but they still lag far behind their mammalian cousins ​​when it comes to rapidly changing movements. lifting environment “.

And yes, they shot a video:

“Under these conditions, you need to use [um sistema de] night vision to avoid accident. For example, walking through a hole can be difficult if the robot cannot see it. While there are some methods of incorporating night vision into leg movement, most are not highly recommended. for flexible robotic systems– said Gabriel Margolis, a graduate student at MIT’s Computer Science and Artificial Intelligence (AI) Laboratory.

To get around this difficulty, Margolis and his team have developed a system that “improves the speed and maneuverability of robots on their feet as they jump over chasms in the terrain.” This system is split into two parts: real-time processing of the footage with a camera mounted on the front of the robot, and another that translates this information into instructions that tell how the robot should move.

In short: the system sees the hole ahead and teaches the platform on which it is installed to jump competently. And the Massachusetts Institute of Technology found it interesting to put all this in a robot named after an animal that runs at speeds up to 130 km / h and strangles prey with neck bites.

Do you know who doesn’t run at 130 km / h? Right: we…

The new method proved to be useful because, unlike other platforms, it does not require preliminary mapping of the area to be covered, which theoretically allows the “cheetah robot that can jump” to move anywhere. The idea is that in the future this platform can be used in response to emergencies in isolated areas or when delivering consumables to hard-to-reach places.

The control of this system is applied artificial intelligence (IA) to process video captured by the camera in real time. According to the Massachusetts Institute of Technology, several “blind” forms (which do not use video) of this application have very robust features, but one difficulty is that the landscape must be continuous without changes in topography. Others depend on the previous mapping, which “trains” the algorithm before actually taking the route.

To get around this, Margolis developed a hybrid format: while the parallel system takes care of capturing video and controlling the joints of the robot, the AI ​​focuses on processing incoming images and “learning” to avoid obstacles when using all directions – not only on the sides and diagonally (implying bypass only when walking), but also vertically (which includes the ability to jump).

To test the effectiveness of the project, the researchers ran simulations based on different types of terrain, through trial and error, setting the evaluation metric as “reward.” The system quickly figured out which actions would bring you the most points and started accepting them by default.

Think of it this way: when you teach a pet to sit down, he takes pleasure. If he sit and roll, two snacks. Etc. But instead of your mongrel, this is a robot with a neural network in its head.

And this neural network made the “cheetah robot that learned to jump” successful in 90% of the tests, outperforming unitary control systems.

“A new feature of our system is that it corrects the gait of the robot,” said Margolis. “If a person is trying to jump over an abyss that is too wide, he can start with a running start to pick up speed, and then put both legs together to make a very powerful jump. Likewise, our robot can customize time and the length of time your feet are in contact with the ground to better traverse the terrain. “

According to Margolis, the current project has only demonstrated that the system works, and it is (presumably) still a long way from practical application. In the future, the MIT team intends to use an internal computer. with more processing power to transmit these readings directly through the platform. In addition, they want to improve the control of the minority so that it does not lose efficiency at different light intensities.

We watched our new videos on Youtube? Subscribe to our channel!