Developers from Google, together with engineers from MIT, Columbia and Princeton Universities, have created a TossingBot robotic mechanical arm that can grab randomly selected small objects and throw them into a container.
The authors of the project say that they had to make a lot of effort to create a robot. With the help of a special manipulator, he can not only grab random objects, but also accurately throw them into containers. It is noted that the choice of subject imposes certain difficulties on the implementation of further actions. Before throwing, the mechanism must evaluate the shape of the object and its weight. After these operations, the decision is converted into action, as a result of which the captured object is sent to the container. The researchers wanted the TossingBot to throw things like a normal person would.
The resulting mechanism visually resembles the robotic arms used on car assembly lines. In action, the robot is able to bend its arm, take one of the objects out of the box, evaluate its weight and shape, and transfer it to one of the compartments of the container, which is designated as a target. To achieve the desired result, the developers taught TossingBot to scan objects, determine their properties, select an object randomly and then capture the target. Machine learning was then applied so that, based on the collected data, the mechanized arm could determine with what force and along what trajectory the object should be thrown.
Testing has shown that the robot manages to grab an object in 87% of cases, while the accuracy of subsequent throws is 85%. Notably, the engineers were unable to replicate TossingBot's accuracy by throwing items into a container on their own.
Source: 3dnews.ru