Knowledge representation and reasoning under uncertainty and robotic applications

Aim:

The aim of our project is the development of a hybrid knowledge base and inference mechanism working with incomplete and uncertain information. Such a representation tool will be used for a design of an incomplete and uncertain model of the relevant part of reality of a cognitive robotic system. The proposed hybrid knowledge base will consist of

(1) graded literal implication of the form: (h₁ ∨ ... ∨ h_n ← b₁ & ... & b_m, c), where h_i, b_j are literals and c is a truth degree of the implication; and of

(2) explicit database atoms with associated hidden Markov models.

By rules of the type (1), we shall represent general common-sense knowledge, while by atoms of the type (2) we shall model training data and estimate unknown testing ones, which will be a suitable means of learning skills.As to the robotic applications we firstly implement our representation tool into a robotic mouse constructed at the University of Trencin, see the following figures (click for details):

front-side view front view side view bottom view

This mouse is able to travel in mazes of the grid form and search the shortest and fastest paths between given points. Its perception is based on an ultrasound sonar system and the inner model of a maze is simply stored in a matrix, which causes the limitation of maze form only on a grid. Using our representation tool, we will be able to extend exploring and searching capabilities of the mouse to an arbitrary surface maze. We also construct a country model of robotic explorer, which will be able to overcome easier unevennesses in the field. Its perception subsystem will be based on a (infrared) digital camera supported by a sonar system for measurement of distances. By the proposed representation tool, we implement the cognitiveness into the explorer. To both the robotic systems, we supplement a voice control unit based on speech recognition techniques.

Methods:

In the development of a hybrid representation tool, we shall use techniques of disjunctive and multivalued logic programming: especially bottom-up computations based on the hyperresolution operator Cp, the query-answering procedure based on ULSLD resolution, and their suitable generalisations. The model semantics will be proposed by means of characteristic fuzzy disjunctive L-models. Learning models from data and speech recognition will developed on the basis of hidden Markov models. In sonar and camera vision, we recognise several levels of processing. In preprocessing, we will apply the standard techniques as in/de-creasing of contrast, focusing, blurring, segmentation. The classification of objects will be proposed by fuzzy logic methods.

Plan:

At the first stage:

we generalise fuzzy disjunctive logic programming to residuated lattices (Guller);
implement a speech recognition and language understanding unit into the robotic mouse (Nagy);
propose a hybrid representation and inference framework as roughly described above (Nagy, Guller);
design and construct the chassis and propulsion (step motors) of the country explorer (Ceska, Janosik, Zervan, Guller);
develop the software and connect the hardware for vision preprocessing on the platform of DSP processors (Ceska, Zervan, Guller).

At the second stage:

we implement the hybrid representation tool into the robotic mouse (Zervan, Nagy, Guller);
design and construct electronic control and power systems for the country explorer (Ceska, Janosik);
propose the classification of objects on the basis of fuzzy logics and develop the software (Ceska, Zervan, Guller);

At the third stage:

we implement the hybrid representation tool into the explorer (Zervan, Nagy, Guller);
implement a speech recognition and language understanding unit into the explorer (Nagy);
mount the vision and sonar system on the explorer (Ceska, Janosik); make the calibration of vision system (Ceska, Zervan, Guller);
make tests in the field (Ceska, Janosik, Nagy, Zervan, Guller).