1. Pavel Nahodil, Michal Petrus, Václav Svatoš
Probabilistic aspects of behavioural control
Department of Cybernetics, Faculty of Electrical Engineering,
Czech Technical University in Prague
Technická 2, Prague, Czech Republic
Pavel Nahodil, Michal Petrus, Václav Svatoš
INTRODUCTION
BEHAVIOURAL INFLUENCE
SIGNIFICANCE
Many tasks in the robot control schedule some navigation in unarranged environment. Their intelligence is independent on sophisticated understanding. The reactive control demonstrates such solution of the problem in the bottom-up way.
We deal in the taxe movement case. Our approach also utilise behavioural features of control system (see Fig.1). These ones are inspirated by insect seeking behaviour using gradient attributes of luminescence, pheromone, or another sign substance.
The behaviour attributes makes another quality in the temporal evaluation of any instant situation. We talk about internal motivations that play excitating and suppresive role at once. The internal rules stimulate partial sub-goals to its satisfying.
The experiments were evidenced a relevance of the behaviour-based method (space cover task) and gradient one (phototaxis) in the robot control. The methods use environment as actual representant of the world without internal map. Therefore, the application depends on the initial arrangenment of the workspace namely.
The internal motivations are integrated from history of actual actions and internal sub-goals (system experience) including explicit motivations (system setup). The actions are encoded to the same basal ways as the sensor representants in this concept.
Benefit of behavioural stimulation
improved orientation to goal at long time horizon
atractor and repelor evaluation links from input to motivation
Benefit of hototaxis
reduces number of the considerous states
reflects only instant relations to significant goals in the environment
allows to fast react to dynamic changes
Benefit of probabilistic processing
probabilistic measure of affect input into motivation
eliminates a low sensitivity caused by small diversity of input signal
the selection mechanism operates with higher uncertainty
Fig. 1 Interface definition of control system T
ACTION COMMAND
NOTICE
CONTROL SYSTEM
LUMINESCENCE SENSOR
ENVIRONMENT
STEP MOTOR
GUIDE
IR PROXIMITY SENSOR
EXISTENCE
INTENSITY
The advanced behavioural synthesis also shows an expert capability of the relevance weighting by its possibility.
Frontal movement -1
Side
movement 1
Fig. 3 An action mapping example
Forward
Backward
Right
Left
Corresponding homogenous areas cover a map of the basic actions (see Fig. 3). The rating is similar to fuzzy grade of desiarability of action.
Behavioural part
activity appetence and motivation
behavioral estimation
Reactive part
environmental feedback
critical actions providing
The system includes:
The distribution of fields in the map of the basic actions depends on the practice or an improved adaptation. We prefere a reinforcement learning to any activity cluster.
The gradient prereqisity was approved in our experiments as well. The behavioural trends does not performed in the brief experiments yet.
SUITABLE METHODS
SELECTION MECHANISM
RELATED METHODS
Our alternative decision omits a knowledge synthesis based on the geometric interpretation of the workspace. We prefer an instant perception of the local neighbourhood [1]. Such processing is based on a functional decomposition that reflect a specific occasion.
Singular inputs and internal stimuli are competed in the following order flow:
The proposed approach could be itemized in the comparation with above-considered methods as follows:
Fig. 4. Schema of the behavioural control system
Fuzzy command fusion
comprehensive design of rules
multi-valued input
rules to compositon of basic axction
available only for continuous integrity inputs
Probabilistic decision
powerful reduction of computational complexity
uncertainty messure injection
Global potential field
reliability in boundary workspace
requirements of known location of objects and their importance
less learning capabilities
Deliberative product planning
complex behaviour realization
time consument and computation complexity
requirements of explicit knowledge about envioronment
Behavioural set adaptation
learning in process of activity
uncertainty activity injection +
This approach utilizes following domains: +
Signal standartization
unification of various sort of input signal
input range equalizer
superposition principle
Probabilistic model of the reactive base
evaluation of input data reliability
quality messure of input data
believness messure
Behavioural motivations
collecting of external stimuli into motivations
integration of motivation tendences
evaluation of action reliability
Fuzzy-based command fussion
input data fashion
actions evolving + +
The reactive rule was yet encoded by the reactictive representation to potential field. Now, we select the most preferred action throug assignment of extreme in the field.
Proabilistic aspects of selection mechanism
fuzzy assignement between input and action
improving quality and sensitivity in the input channel
exact decision +
The probabilistic estimation works in the probabilistic field, which is construed at the same level as the potential field. +
REACTIVE REPRESENTATION
EXPERIMENTS
REFERENCES
We process two principal sort of robot incoming information:
Existence
pure reactive stimuli
binary values (infra-red proximity detection)
Intensity
advanced knowledge for reactive decision
continuous range of input values
Proposed system architecture was tested on the mobile robot platform[3]. There were achieved experiments focused on gradient-based navigation in unknown enviroment.
Implemented robot’s behaviour embodies reactive character. The executed actions was educed from local environmental information without design of internal representation.
[1] Brooks, R. A.: Intelligence without Representation. In: Artificial Intelligence, Vol. 47, pp MIT Press, 1991
[2] Starý, J.: Rozšířený HW pro mobilního robota. Diploma work, CTU Prague In czech.
[3] Kurzveil, J - Maixner, V - Svatoš, V.: Autonomous Mobile Robot Platform. In: Proceeding of 3th International Student Conference on Electrical Engineering "Poster 1999", p. IC 26, Prague, May 1999.
[4] Khatib, O.: Real-time Obstacle Avoidance for Manipulators and Mobile Robots. The International Journal of Robotic Research, Vol.5 No.1, pp.90-98, 1986.
[5] Balch, T.: Behavioral Diversity in learning Robot Teams. Doctoral thesis, Georgia Institute of Technology, December 1998.
[6] Large, E. W. - Christensen, H. I. - Bajcsy, R.: Scaling the Dynamic Approach to Path Planning and Control: Competition among Behavioural Constraints. The International Journal of Robotic Research, Vol.18 No.1, pp.37-58, 1999.
These factors take a potential field representation [4] by its goal rate:
Attractor representant
goal reinforcement stimuli
motivation rate
Repelor representant
conflicts against goal
suppresive rate (see Fig.2)
Fig. 5. The robot trace in the light labyrint
Fig. 2 Repelors in frontal detection case
Rate
Frontal movement
Side movement -1 1
lunminescence sensor (light intensity related to movement rate) [2]
infra-red bumpers (existence to behavioural reduction)
beacon extetion (guide notice/alert)
Robot equipement
This research has been supported by the Ministry of Education, Youth and Sports of Czech Republic. GRANT FRVŠ No