There are some tasks for which traditional robots—rigid and metal—are simply not suited. WITH
MIT researchersThe institute has developed a special algorithm to solve this problem. It will help engineers develop software robots that collect more useful information about the environment. The deep learning algorithm suggests optimized placement of sensors in the robot's body. This, in turn, allows it to better interact with the environment and perform assigned tasks. “The system not only learns a specific problem, but also how best to design a robot to solve that problem,” explains Alexander Amini of MIT.
The research will be presented at the AprilIEEE International Conference on Soft Robotics. Co-lead authors are Alexander Amini and Andrew Spielberg, graduate students at MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL). Other co-authors include MIT graduate student Lilian Chin and professors Wojciech Matusik and Daniela Rus.
Soft-bodied robots are flexible and malleable - theylook more like a bouncy ball than a bowling ball. Their main problem is that they are infinitely dimensional. Any point of a robot with a soft body can theoretically deform in any way possible. This makes it difficult to create a soft robot that can display the location of its body parts. Past attempts have used an external camera to determine the position of the robot and send this information back to the robot's control program. But the researchers wanted to create a soft robot that didn’t depend on outside help.
“You cannot accommodate an infinite number ofsensors on the robot itself, - emphasizes Spielberg. "So the question is, how many sensors do you have and where do you place them to get the most out of your investment?"
The team turned to deep learning for the answer.
Researchers have developed a new architecturea neural network that optimizes sensor placement and learns to perform tasks efficiently. First, the researchers divided the robot's body into regions—"body parts." The deformation rate of each particle was entered into the neural network. Through trial and error, the network learns the most efficient sequence of movements to perform tasks, such as grasping objects of different sizes. At the same time, the network keeps track of which parts are used most often and selects less used ones from the input data set for subsequent network testing.
By optimizing the most important parts of the robot's body,the network also suggests where to place the sensors on the robot to ensure efficient operation. For example, in a simulated robot with a grasping arm, an algorithm might suggest that sensors be concentrated in and around the fingers, where precisely controlled interactions with the environment are vital to the robot's ability to manipulate objects. While this may seem obvious, it turned out that the algorithm far surpassed human intuition about where to place the sensors.
The researchers compared their algorithmwith a number of expert forecasts. For three different soft robot designs, the team asked roboticists to manually select where sensors should be placed to ensure tasks such as grasping various objects could be carried out efficiently. They then ran simulations comparing touchscreen robots to touchscreen robots. And the results were not close. “Our model significantly outperformed humans on every task. Although I was sure that I knew where to place the sensors… - concludes Amini. “It turns out there are a lot more subtleties to this problem than we originally expected.”
Read more
Physicists have created an analogue of a black hole and confirmed Hawking's theory. Where it leads?
Scientists have discovered the mythical particle of Odderon
The most mysterious natural phenomenon. Where does ball lightning come from and how is it dangerous?