Mobile robot control by lerned behaviour.

Pozna, Claudiu ; Alexandru, Catalin

Abstract: The aim of this paper is to present a general view about the autonomous driving researches made in the University of Applied Science Heilbronn, Germany. More exactly we will present aspects about our autonomous car control system design. This will be the occasion to focus on the tactical level of the mentioned control system in order to present the behavior based control strategies. One of the first results of the mentioned strategy is the trajectory tracking. Key words: autonomous car; driving robot; control system architecture, trajectory tracking.

1. INTRODUCTION

Nowadays, autonomous car designs are taken into consideration more and more. The mentioned concept can be defined like a car which is able to drive itself. So this means that we deal with a car which copies a human driver's performances. Is this a mobile robot? The answer is yes, if we consider only the navigation performance, but no, if we have in mind that it is necessary to add specific interfaces that are not needed in the mobile robots case. The present work has ignored this difference and has focused on the driving (navigation) performance that means that our autonomous car (Pozna & Troester 2007) is in fact a mobile robot.

This paper presents aspects about the ACC autonomous car that we have constructed in the University of Applied Science in Heilbronn, Germany. The presentation will focus on some original elements which refer to the control architecture of the mobile robot. More precisely, we will present our three level control system which operates with a collection of programs named behaviors.

Our control system architecture is based on the human driver behavior model concept. So, in order to present our ideas, some preliminary discussions are necessary. We know that the "Driver's Behaviors" model is used in the simulation (Al-Shihabi 2001; Quispel 2002), and also in autonomous car design field (Bengtsson 2001). The first researches on the subject started in 1950 and began with the "Skill-based driving model", continued with the "Motivational model" which considered the drivers' emotional state (from this class we can enumerate the "Risk compensation"; "Risk avoidance" and "Risk threshold models"). Recently, the model turned to a "Hierarchical control structure". The "Hierarchical control structure" (Milchon structure) divides driving into three levels of control: a strategic level, which establishes the goal of the driving, a tactical level, which finds the solution to accomplish the goal and an operational level, which implements this solution on low level control of the vehicle. Behind this "Hierarchical control structure" many scientific papers consider and develop problems like: "Longitudinal behaviors models" (Bengtsson 2001); "Lateral behaviors model" (Ungoren 2005); etc. The solutions of these problems are varied: "Linear optimal Control", "Heuristic human driver models", "Adaptive control strategy", "Neuronal Network and fuzzy logic", "Mental models", etc.

Because we have intended to make a heuristic approach, we have been interested to find control programs architectures which model the human behavior. Such architectures are presented in (Al-Shihabi 2001 and Quispel 2002).

Some conclusions about these briefly overviews: --In the scientific literature referring to "Driver Behavior Model" we have found several results which can be adapted and used in our mobile robot control;

--Recent works accept the Milchon three levels architecture; Many papers focus on the tactical level where the program must find the solutions in condition of changeable driving circumstance.

2. THE CONTROL ARHITECTURE CONSTRUCTION

Our idea starts from this point: we consider that it is more suitable to model and implement the "human driver decisions act", than the "human driver actions". This idea transfers the approximation of the human driver behaviors from a mechanical to an artificial intelligence problem. This kind of problem involves preliminary analyze which must answer to the following questions: "What are driving behaviors?"; "Can we obtain some fundamental true about these behaviors and use them in our construction?" We have made a phenomenological research which starts with the semantic characteristics of "Driving behavior". First, it is important to establish the category tree of this word: {act [right arrow] activity [right arrow] (behavior, practice, ...)}. According to this, the behavior is: "an action or a set of actions performed by a person under specified circumstances that reveal some skill, knowledge or attitude". From the scientific literature which concerns the driving behavior (Huang 2000; Liu 2001; Salvucci 2005) and from our experience, the driving behavior has a special feature. To describe it, we focus on the word "custom" which belongs to the same category tree {act [right arrow] activity [right arrow] practice [right arrow] habit,.} and which is defined as: "accepted or habitual practice". In many situations, these habits have a special nature: automatism--any reaction that occurs automatically without conscious thought or reflection. Using the previous definition, "Driving Behavior" is an action or a set of actions performed by a person under driving circumstances, actions which tend to be transformed in habits and even in automatisms. In fact the "Driving Behavior" is composed of a series of behaviors (the driver's behavior when he makes the ignition, the driver's behavior when he stops the car, ... etc.). From the mentioned theoretical and practical research, we established the following "fundamental truth" for the "Driving Behavior":

1. A priori, the driver establishes the current driving goal;

2. A behavior is a set of actions;

3. These behaviors are linked together, creating a system which allows the driver to obtain solutions in the driving circumstance;

4. The translation from one behavior to another is triggered by the occurrence of an event;

5. This system is developed by learning--experience;

6. Behaviors presume decisions with an incomplete set of information;

7. In time, these sets of actions tend to be transformed in habits and automatisms.

These propositions agree with the well-known three level architecture of Milchon: the strategic level, where the driver establishes his goal, the tactical level, where the driver finds the solution to accomplish the goal and the operational level, where the driver implements these solutions. Using these propositions, we can focus on the tactical level and model (approximate) the "Human Driving Behaviors" by a collection of high linked programs (behaviors) which are stored in a memory. The decision to run a certain program is made by a manager program. This decision is based on the driving goal and on the environment understanding (driving circumstance). Each program (behavior) is a set of instructions (actions) which impose parameters and trigger actuators. Using these seven propositions, we can imagine the utility of state machines, for handling the behaviors, fuzzy logic to enable decisions or to describe the environment and neuronal network to implement the learning processes.

The aim of figure 1 is to make our concept more understandable and to allow the necessary explanations:

--The strategic level, where the robot receives its task (goal) is an interface which helps the human operator to impose the goal;

--The "Program Manager" analyzes the goal versus the driving circumstances which are obtained from the sensors; the result of this process is the status vector of the robot (the desired position, velocity, etc.) and also the decision to run a certain program from the "Behaviors" subsystem;

--The "Behaviors" contains three parts:

--The "Error Machine" which compares the status vector with state vector (the positions, velocity, etc. obtained from the sensors);

--The "Behavior Programs": is a collection of programs (behaviors); each program is able to solve a special environment situation (ignition, emergency stop, zero position, errors....);

--The "Actuators Manager" which manage the actuators of the robot;

--The "Output Interface" allows the states and the errors reading and also the robot state history memorization; The "Actuators Communications" outputs data to the microcontrollers of each actuator;

[FIGURE 1 OMITTED]

In order to build the "Behaviors" subsystem, it is important to imagine the structure of the programs which are included in the "Behaviours Programs" (see figure 2). There are different types of such programs: "basic behaviours", "error behaviours" and "simple behaviours". The main differences between these programs are the connection type and also the direction of information flow.

[FIGURE 2 OMITTED]

The control program has been build with Matlab (xPC toolbox). The navigation problem that we intend to solve can be defined in the following way: the task of the mobile robot is to follow a certain trajectory between two points; the trajectory is mathematical defined on a map, but (initial) unexpected obstacles must be avoided during the navigation. In order to be able to accomplish this goal the "Behaviors Program" (see figure 1) must contain the following programs (behaviors): start the car; follow an a priori defined trajectory; return to the a priori defined trajectory; avoid the obstacle and return to the a priori defined trajectory; stop de car. Experiments have been made to track linear or circular trajectories.

3. CONCLUSIONS

We have presented aspects about the control system design. The original achievements of these works have been highlighted. More precisely, original control system architecture, based on human behavior model was proposed. We must mention that nowadays, only a part of the control architecture is made, so future work will develop this architecture. We intend also to develop our environment sensory system for circumstance recognition, and "Program manager" development.

4. REFERENCES

Al-Shihabi, T.& Mourant, R. (2001). A Framework for Modeling Human-like Driving Behaviors for Autonomous Vehicle in Driving Simulators. Proceedings 5th International Conference on Autonomous Agents, June 2001 pp. 286-291.

Bengtsson, J. (2001). Adaptive Cruise Control and Driver Modeling, Department of Automatic Control Lund Institute of Technology Lund, November.

Huang, S. (2000), Design and Performance Evaluation of Mixed Manual and Automated Control Traffic. IEEE Transaction on Systems Man and Cybernetics pp. 127-136.

Liu, A. (2001). Modeling and Prediction of Human Driver Behavior. Proc. of the 9th International Conference on Human Computer Interaction, Aug, 2001, pp. 235-244,

Pozna, C. & Troester, F. (2007) Research on the ACC Autonomous Car. Journal of Automation, Mobile Robotics & Intelligent Systems Vol.1, No.1, pp32-40

Quispel, L. (2002). Automan, a psychologically based model of human driver. Experimental and Work Pscychology. Department of Psychology, University of Groningen Available from: http://www.tcw2.ppsw.rng.nl, Accessed: 2005-08-08.

Salvucci, D. (2005). Modeling Driver Behavior in a Cognitive Architecture. In Press, Human Factors February 16.

Ungoren, A. (2005). An Adaptive Lateral Preview Driver Model. Vehicle System Dynamic.vol42, pp 225-259.