The American scientific research team trained an intelligent manipulator that can recognize its body independently by using artificial intelligence algorithm. Without input of knowledge of physics, geometry and kinematics, this arm, which is the same size as human arm, can recognize its shape in a short time and make a higher level of motion decision.
In the new study, Columbia University team first allowed a manipulator to move at will and collected about 1,000 trajectories, each containing 100 moving nodes, then used in-depth learning technology to build a self-model of the manipulator.
At first, the manipulator did not know how to shape itself and how to connect the joints. But after less than 35 hours of training, the error between the self-model and the actual physical shape of the manipulator was very small, and the manipulator could accomplish the task more accurately.
Experiments show that under the “calibrable” mode, which allows the manipulator to adjust itself according to the trajectory of motion, it can clamp several small balls into the cup with a 100% success rate. Researchers also replaced the manipulator with a 3D printed defective part to simulate the condition of its body damage.
The results show that the self-model of intelligent manipulator can discover the changes of “body”, and after retraining, it can complete the experiment task at the same level as before.