Abstract

Robust and Adaptive World Models for Mobile Robots

IEE Seminar on "Self-learning robots"
January '96 - Savoy Place, London, UK

1. Motivation

This seminar is especially dedicated to adaptive spatial modelling for mobile robots. As the central optimization criterion, the navigation performance, i.e. the performance of moving is employed. Criteria like the flexibility of the user-interface or the performance of a planner for manipulations, etc. are not considered. Due to the fact that the navigator uses the internal relative position, the self-localization task is included in the model as well as in its adaptation process. Finally the exploration strategy in previously unknown environments influences significantly the methods of world model adaptation. A mobile platform in unknown environments, not prepared in the form of defined absolute positions (landmarks) is considered in the following. Furthermore it is assumed, that the robot's sensor equipment is adequate for the working environments, i.e. any feature meaningful for manoeuvring, for the application, or concerning the robot's security aspects can be detected. Then the elementary tasks, essential for the considered mobile platform (and for most mobile robots) can be summarized:

Exploration
This task concerns the problem of gathering information from the actual working environment efficiently, due to some optimization criteria.
Spatial Modelling
The information sampled from the working environment has to be integrated as a spatial representation, applicable for a navigation task.
Self-Localization
Based on the spatial model, the robot has to preserve control over its position and orientation in relation to the current environment. The solution of this task states an essential precondition for the navigation task.
Navigation
Employing the spatial model together with the knowledge about the robot's relative position, the task of finding adequate manoeuvres, in order to reach a given goal position, has to be solved.

2. Qualitative and Adaptive World Modelling

The central motivation of qualitative topological world models (QT-Models) is the basic mobile robot task: "Recognize places you have seen before!". In this seminar this task will be approximated by extracting "situations" (i.e. recognized places) together with their topological neighbourhood from the current sequence of sensor-samples, rather than modelling the boundaries of the detected obstacles and objects in a metric manner. Assuming a stable situation-recognition-process and a technique for moving between distinct situations, the concept of a qualitative, topological world model suggests a human-motivated basis for a navigation. The main concept has already been proposed by Kuipers et al. [5], but here the construction process was carried out using explicit rules, not statistical techniques. Therefore the real-world abilities of the Kuipers approach are, in the opinion of the author, limited. The world model proposed in this seminar is based on clustering techniques introduced by Kohonen ("self-organizing-maps", [4]) and Fritzke ("growing cell structures", [1]) together with some previously proposed extensions by this research group [3]. Due to a couple of specific autonomous robots-constraints, these structures are modified to cope with realtime-aspects, lifelong learning, "local forgetting", and correlation. The requirements regarding the employed computational power as well as the sensory equipment are kept at a very low level. The processes coping with navigation, track control, self-localization and adaptive world modelling are handled in realtime (while driving at 25 cm/s) on one (CISC-) CPU. The sensory equipment is limited intentionally to binary whiskers, passive light sensors, and rough odometry. Based on these assumptions and restrictions the robustness of the techniques can be shown in experiments performed with the real mobile platform. The technical details of the proposed methods as well as experimental results and discussions regarding reliability, robustness, consistency and practical relevance can be found in [7] and [8].

3. Conclusion

Beside the previously published features of the qualitative, topological world modelling techniques, some criteria (or guidelines) for the choice of the employed world model were introduced in this seminar. One aspect is reliability, for example in a non-error-tolerant environment. If that is to be a central aspect of the robot-task, an exact model may be required to be able to plan safe paths. Similarly, if a guarantee of accuracy when following a path is needed, the exact geometric information may be necessary, i.e. the proposed qualitative modelling is not sufficient. On the other hand, if the main focus is on simplicity, stability or qualitative aspects of the task, the qualitative topological map techniques may be the first choice. Especially the small requirements for computational effort and sensor equipment together with a high degree of robustness is an unique feature. The experiments have shown real world abilities offering sufficient information for navigation purposes as well as a stable self-localization method.

References

[1] Bernd Fritzke
Growing Cell Structures - A Self-organizing Network for Unsupervised and Supervised Learning
Technical Report 93-026, International Computer Science Institute, Berkeley, California
[2] Bärbel Herrnberger, Uwe R. Zimmer
Deriving Baseline Detection Algorithms from Verbal Descriptions
Artificial Neural Networks and Expert Systems '95, November 20-23, 1995, Dunedin, New Zealand
[3] Herman Keuchel, Ewald von Puttkamer, Uwe R. Zimmer
SPIN - Learning and Forgetting Surface Classifications with Dynamic Neural Networks
Proceedings of the ICANN '93, Amsterdam, The Netherlands
[4] Teuvo Kohonen
Statistical Pattern Recognition Revisited
Advanced Neural Computers / R. Eckmiller (Editor), Elsevier Science Publishers B.V. (North-Holland), 1990
[5] Benjamin J. Kuipers, Yung-Tai Byun
A Robust, Qualitative Method for Robot Spatial Learning
Proceedings of the AAAI 1988
[6] Jun Tani, Naohiro Fukumura
Learning Goal-Directed Sensory-Based Navigation of a Mobile Robot
Neural Networks, Vol. 7, No. 3, pp. 553-563, 1994
[7] Uwe R. Zimmer
Robust World-Modelling and Navigation in a Real World
Neurocomputing 1996
[8] Uwe R. Zimmer
Self-Localization in Dynamic Environments
IEEE/SOFT International Workshop BIES'95, May 30 - 31, 1995, Tokyo, Japan

(figure 1 : Mobile robot 'ALICE')