Robots learning to throw stuff

A person does nothing but throws things. But the actual throwing action is a fairly advanced move. Google announced that TossingBot is a robot that can learn to throw things like this through machine learning.

Throwing stuff is a fairly complicated move. Because the throwing object state, the timing to put the object, and the physical action after throwing are intertwined. Throwing the same object can allow the robot to handle it correctly, but it is quite difficult to program a robot from the beginning by picking randomly entered table tennis balls, fruits, tools, and so on.

Google has partnered with Princeton University, Columbia University, and the MIT research team to develop a robot that can learn to throw things through machine learning. The TOC Shinbot recognizes the objects you want to visually throw and you can repeat what you actually throw objects to learn what happens when you throw an object and how you can get the results you want.

When putting the ping-pong ball into the box with the robot arm, set the position to match the ping-pong ball trajectory. A completely different banana model is likewise thrown, allowing the box to fine-tune its position and capture it. The act of throwing things as you see in throwing garbage in garbage cans or throwing naked clothes in a laundry basket is a combination of various factors. You have to pick up the object correctly and shake your arm at the proper speed and remove your finger at the right time. Depending on the throwing object, when you throw it, it will change its orbit as well. Since humans throw a variety of objects from their young age, they learn these behaviors naturally.

Throwing objects can be done by using only a part of the body without using too much energy and moving objects to places where they are physically inaccessible. Throwing more than moving things one by one is also as efficient as moving goods quickly. If the robotic arm is able to throw things properly, it can be used at delivery points, factories, and disaster sites.

Toe Shinbot learns the structure of throwing objects by picking up objects from a box containing various objects and repeating the actual throwing action. Repeat the action several times to accumulate data such as how to shape the object, shake it, and when to put the object.

Just as human beings throw things, Tohkingbot determines the shape of objects by visual elements before throwing them. At first, it fails to catch things, but eventually catches them correctly. Of course, it repeats learning in throwing motion. The TOCINBOOT can repeat 10,000 attempts over 14 hours, and after 10,000 hours of learning, it can hold the object at 87% accuracy and throw the object at the proper position with 85% accuracy.

The TOC Shinbot learns through simple pre-set physical information and machine learning. By providing models in advance of how the world moves in physics, learning by machine learning is also facilitated. In fact, Toki Shinbot, who has learned a lot, has been able to throw objects more accurately than humans. For more information, please click here .