"Because neural networks and sticks": how unusual robot learn to walk

Today, robots of different shapes is very difficult to surprise. However, a group of scientists from Japan seems to have succeeded. They built a strange robots of the sticks ... and other materials at hand. Moreover, thanks to the highly advanced artificial intelligence system built on the basis of neural networks, the robot learned to walk. And as strange as it may sound, but this approach is very promising future.

How to create a robot

Typically, the design of robots, scientists need to have a very clear idea of ​​what the robot should do, how it will move and what actions it will need to perform. In this case, you first need to build a prototype of the future devices to make sure that all their calculations were correct and to identify design flaws. But in this development process is not finished. After the formation of understanding of how the model will work, you will need to "teach" it to those actions that it must perform, whether working as a paddle, walk or move in other ways.

The robots of the branches

But, as scientists from the University of Tokyo and the company's Preferred Networks declare the creation of a robotics process does not necessarily have to be so complicated. Back in December last year, they began to experiment on creating robots based on the servo and readily available materials (eg tree branches).

Read also: Robots learn to recognize objects by using sight and touch

In this case, even before the creation of model robots are studying the movement in a special computer simulation, where they were taught to walk an advanced neural network based on deep machine learning. To download the data in a neural network model, the scientists simply take a few branches of a suitable size and the robot body. All this is subject to 3D-scanning. The resulting models appear in a special computer program which "collects" them together and begin the process of learning.

During these "lessons" given by the model of the future behavior of the robot. For true motion neural network "rewards" future mechanism, but for the wrong - "punishing". In addition, it is possible to manually adjust the movements, as well as the system that provokes the emergence of random events during the training in order to maximize the robot was ready for what awaits him in the real world. When the training is over, can only assemble the model and load it in the behavior of the algorithm.

Robots controlled system built based on Arduino Mega and servo Kondo KRS-2572HV

Why create such robots?

In fact, the potential of such a system is enormous. In fact, all you need - it's "computer and a pair of a motor." The remaining parts can be assembled directly at his feet. Such robots can be used for reconnaissance in areas where a person does not get, and send complete mechanisms are too expensive. In addition, these robots can be collected, such as the colonization of other planets. You can, for example, send the assembly module with a margin of microprocessors and actuators, and other items already collected on arrival at the mission.

Discuss this and other news you can in our in the chat Telegram.