Hard to be a robot when one is faced with the real world! By default, a robot is capable of performing tasks for which it has been programmed, even gestures that seem the most basic for a human.
Researchers at the Carnegie Mellon University Robotics lab has a robot Baxter of a Deep Learning algorithm. The objective is to allow the robot to learn himself a simple task using its artificial intelligence. A promising approach, but oh how much laborious.
Deep Learning empowers the robot to learn from its mistakes
Nobody needs to pay the slightest attention to enter a remote, scissors, or a toy on a table. So that a robot can perform such a gesture without programming and without dropping the object, it is a different matter. How to catch those objects to forms the most different without them falling?, not so simple. U.S. researchers have tried to see if a Deep Learning algorithm may allow the robot to learn by him even, through trial/error. They therefore instructed their algorithm in the ‘head’ of the robot, placed it in front of a table full of small objects and energized Baxter. The Baxter is driven for more than 700 hours during which he made 50,000 tests on 150 different objects. Lifeline learning phase for its algorithm that includes 18 million of parameters which has proved to be effective. Faced with an object, the robot is shown capable of predicting how to take without dropping it with a probability of 80%. In relying on simple rules as catch the object at its Center, on a not too narrow area, the probability of success is only 62%.
Researchers have shown that deep learning will enable robots to learn to interact with their environment, learn like children… but what effort!
Translation : Bing Translator
“Deep-Learning Robot Takes 10 Days to Teach Itself to Grasp”, MIT Technology Review, October 5, 2015
“Supersizing Self-Supervision: Learning to Grasp from 50K Tries and 700 Robot Hours”, Cornell University Library [pdf], September 23, 2015