* ROBOS is preliminary project name.

The world will soon be populated by hundreds of thousands of robots of various "professions". Today, robots 1.0 are rather primitive machines that execute the program with a sequence of actions and a set of "if-then" rules. Such robots require many hundreds of hours of software programming and tuning. They can't be quickly trained to other application and their application is limited to monotonous routine operations.

The future robot 2.0 is a self-learning machine that will learn as a child and "be excited" with its results, trying to learn better and better. Robot training will be performed completely without software programming on real objects or in the virtual world. This robot is actually a next-generation computer, which will require a new operating system and new "programming", where the robot's brain will be the assembly of various neural networks, created and trained by various players of the ROBOS ecosystem.

Once such open operating system for robots is created, each manufacturer will be able use it and to focus on his specifics, not spending resources for R&D in the area of AI,robot's adaptive behavior and education.

Such approach becomes real only if the robot is able to learn in real time. Today's deep learning algorithms based on the "back propagation" method can not provide this. Most of current algorithms have computational complexity quadratically dependenton the number of parameters and require supercomputer computing power for network training.

ROBOS will use a new model for training neural networks, developed by one of the project founders. It leverages the technique of "deep structural learning", based on the discrete sparse coding scheme for any data. The computational complexity of this learning algorithm depends linearlyon the number of parameters. For example, on tasks for processing unstructured texts, this modellearns a thousand times fasterthan existing algorithms, without requiring a supercomputer and is able to learn in real time.