Projects / Programmes
Environment identification and mobile robot navigation with neural networks
| Code |
Science |
Field |
Subfield |
| 2.07.07 |
Engineering sciences and technologies |
Computer science and informatics |
Intelligent systems - software |
| Code |
Science |
Field |
| P170 |
Natural sciences and mathematics |
Computer science, numerical analysis, systems, control |
neural networks, mobile robots, navigation, control, dynamical systems
Researchers (1)
| no. |
Code |
Name and surname |
Research area |
Role |
Period |
No. of publicationsNo. of publications |
| 1. |
14300 |
PhD Branko Šter |
Computer science and informatics |
Head |
2002 - 2004 |
0 |
Organisations (1)
Abstract
We believe that mobile robots will be a part of everyday's life in the near future. Therefore it is important not to lag behind in the field of autonomous mobile robot navigation. In order to successfully operate in indoor environments, a robot should be able to master basic tasks such as environment exploration, self-localization and navigation (finding a path to a goal).
Our proposed approach consists of the application of recurrent neural networks (RNN) for implicit modeling of the robot's environment. Since the robot is mobile, its sensory space evolves in time, therefore it can be treated as a dynamical system. Learning of the environment corresponds to the identification of a dynamical system. Since the modeling of such a system at a small scale (seconds or less) is practically impossible, we apply certain points (landmarks), at which a decision of the robot for the next action is sensible and possible; otherwise the robot moves according to a certain reactive or low-level program. Such discretized dynamical system can be described in the form of a finite state automaton (FSA), on the level of which the path planning during navigation is conducted.
An essential drawback of the automaton comes to light, when unexpected situations arise as a consequence of the real environment and the real robot (especially sensors). Therefore RNN is applied, which "learns" the structure of the environment, i.e., the corresponding automaton, on the basis of an input-output sequence (actions and sensors) during the exploration phase. In this way a continuous dynamical system (RNN) is applied to modeling a discrete dynamical system (FSA). One of the utilities of RNN is a relative tolerance to noise and errors, which are unavoidable in reality.
In order to significantly increase the autonomy, we plan to develop a method on the basis of some kind of self-organization on the sensory- and perceptual level, which would enable the robot an automatic discovery of landmarks.
Once the robot learns the environment, it can navigate more efficiently, while with a limited success it can navigate even before. During the navigation to a desired goal the robot conducts planning on the simulation-level ("mentally"), where it seeks an appropriate action sequence for the optimal path.
At the same time we plan further research in the field of recurrent neural networks with the goal to increase the robustness and the speed of learning.
