Robots learn by experimentation and observation

Researchers in the European project XPERO have developed a machine learning method that enables a robot in a position to learn fundamental mathematical concepts such as location and orientation in a coordinate system.

The robot initially moves aimlessly through the neighborhood and is characterized, through its sensor data, without being aware of the information contained therein. The algorithm takes these sensor data to generate a model that allows the robot to predict how objects will change their position in his neighborhood as a result of his own movement.

“What is a trivial matter for a man is a robot for a rather difficult problem,” Jure Zabkar, Ivan Bratko shared by the University of Ljubljana, the inventor of the algorithm. “Our robot has less knowledge than a baby. An object seen is meaningless for him. He is just doing color blobs and edges.”

The robot does not know the concept of an object, or a position of an object in a coordinate system, nor does he know how this changed when he moved himself. The machine is therefore not given in advance; he should learn a coordinate system, nor how to get it or what it’s for.

“We have developed a mechanism that allows the robot to extract regularities from the sensor data and to translate them into a model or a theory that enable the robot to better explain and predict what is going on around him right . Learning a coordinate system is simply a manifestation of this ability, “said Zabkar.

What at first more like an academic basis problem excludes, also has an enormous technical relevance, the project coordinator Erwin Prassl explained by the Bonn-Rhein-Sieg. The XPERO-project lays the first foundation stones for a technology that has the potential to be a key technology for the next generation to be of service robots that keep our homes clean, mow our lawns or clean our shoes.

Existing products are intelligence-free, pre-programmed devices. You can only run a single pre-programmed task. Neither are they able to start new, not previously familiar tasks, nor to cope with unforeseen operating conditions. Future service robots must however be able to learn based on their existing knowledge and sensor observations entirely new concepts and models and to comply with this new knowledge to new tasks.