Sensor Data Fusion for Control of Mobile Robots using Fuzzy Logic

Sensor Data Fusion for Control of Mobile Robots using Fuzzy Logic

Copyright © IFAC Intelligent Manufacturing Systems. Bucharest. Romania. 1995 SENSOR DATA FUSION FOR CONTROL OF MOBILE ROBOTS USING FUZZY LOGIC H. Rot...

1MB Sizes 2 Downloads 98 Views

Copyright © IFAC Intelligent Manufacturing Systems. Bucharest. Romania. 1995

SENSOR DATA FUSION FOR CONTROL OF MOBILE ROBOTS USING FUZZY LOGIC H. Roth, K. Schilling

Hochschule for Technik und Sozialwesen, Fachhochschule Ravensburg- Weingarten , Institut for angewandte Forschung, Postfach 1261, D-88241 Weingarten, Germany, e-mail: [email protected]

Abstract: In automatic production processes guided vehicles are a key component for intime material flow. The development of a new generation of mobile transport vehicles has the emphasis to give them the capability of working autonomously over a certain period of time providing increased flexibility . One essential aspect to provide autonomy is to determine a precise model of the environment detected with sensor equipment in the area of navigation, obstacle avoidance, target acquisition and finally high accurate target docking. On the one side these sensors can deliver completely different and partly contrary information about the same surrounding, on the other side the measurement signals are noisy . Additionally, using low-cost sensors the measured values are not very precise. While extended Kalman filtering has proved to be an appropriate classical method for decentralised data processing, in this paper it will be shown especially for collision avoidance, that Fuzzy Logic can be successfully applied for data fusion of different sensors and improvement of noisy signal information, especially in higher levels of mobile robots control hierarchy. Keywords: mobile robots, autonomous vehicles, sensor data fusion, collision avoidance, fuzzy control.

I INTRODUCTION

Currently a second generation of AGV's is under development. providing an increased flexibility. Here the information on the path is stored in the vehicle's board computer. instead of in the wires laid into the floor. This enables a huge flexibility regarding an effective path planning. in particular with respect to collision avoidance and docking to target stations. The technology progress and cost decrease of modern computers enable an efficient reconfiguration of the vehicle's characteristic parameter via software. providing a broad band of adaptation capabilities. It is of benefit for modern flexible production cells and for manufacturing of different product variants .

From an economical point of view it is interesting to substitute a single highly accurate but expensive sensor by several low-cost sensors, which are more inaccurately, if by post-processing a similar accuracy in the measurement results . This graceful degradation by using low-cost measurement systems can be compensated by intelligent sensor systems based on different types of sensors combined with intelligent post processing. In industrial production lines Automated Guided Vehicles (AGV's) are employed to support the flow of materials. Here initially inductive wires placed in the floor provided the reference path and appropriate control systems corrected measured deviations . A major disadvantage of this approach is the static paths determined by the fixed wiring. causing high costs and delays if paths have to be changed.

But to provide the robot with this autonomy capabilities it is necessary to equip it with suitable sensors to recognise its environment. what especially means to detect obstacles and to dock at the final target. According to economical aspects low-cost sensors are favoured in industrial applications. 317

Table I: Data fusion technigues for multi-sensor s:istems Principle stochastic

Fig. I.

Method

Data fusion technigue occupancy Bayes estigrid mation

application obstacle detection, navigation

stochastic

Kalman filtering

filtering on the base of a system model

correction of sensor drift and nOise, fusion of sensor data

information processing

fuzzy logic

logical reasoning

qualitative modelling of problems

information processing

neural nets

trained behaviour based on eXEeriment

classification, obstacle detection

measurement system and the cost. With respect to cost the multi-sensor data fusion makes especially sense, if by combining several low-cost sensors a similar result is achieved at a cheaper price than by using one very expensive sensor system. A typical example for this is in the area of mobile robots the use of sensors for obstacle detection. There are available on the one side expensive but accurate 3-D laser scanner or stereo vision cameras, which demand a high post-processing effort, and on the other side cheap but inaccurate ultrasonic and infrared distance sensors. Combining different ultrasonic sensors with the infrared sensors to achieve a qualitative good measurement will be the topic of the following chapter.

Weingarten's Autonomous Mobile Robot (WALTER) on which the developed data fusion techniques are tested.

The aim of this paper is to study control strategies for collision avoidance. These local strategies are based on relative distance and velocity measurements to other objects achieved with low-cost sensors and post-processed with adequate data fusion tech~iques. It is intended to demonstrate that fuzzy logic IS a suitable technique for sensor data fusion combining several sensors for distance measurement in the mobile robot application. The algorithms are implemented on the mobile robot shown in Fig. I and the results are verified in experiments with this robot. In the following section principles of sensor data fusion are sketched. Section 3 provides further details of the fuzzy control algorithms for collision avoidance based on different sensor systems. Section 4 gives a conclusion.

Classical methods of sensor data fusion (Abidi , et al. . 1992; Kim and Kokar, 1994) like Kalman filtering (Doyle and Roth, 1995) or occupancy grids (Elfes. 1989) are based on the stochastic modelling of the sensors. In situations requiring a very costly modelling or when only qualitative information is available the use of information processing techniques like fuzzy logic or neural nets is . of advantage (Schilling, 1995). This paper is concerning a hierarchy for control of a mobile robot using fuzzy logic for evaluation of different measurements With respect to obstacle detection combining different methods given in Table I .

2 SENSOR-DATA-FUSION

3 COLLISION A VOIDANCE STRATEGIES

In multi-sensor data processing it is the aim to combine the single sensor measurements of different sensors for the same physical quantity and to increase the quality of the resulting sensor information. Of course, the improvement must be reasonable compared with the increasing complexity of the

The base for navigation and localisation of an autonomous mobile robot and also for preventing collisions with fixed or moving obstacles are actual sensor measurements combined with a knowledge base representing characteristic landmarks of the 318

controls. The fuzzy sets are defined by the different relative position classes of objects, characterised by angle and distance. The inference rules connect to each fuzzy set a related control for the robot's speed and its steering angle. In the real hardware also these output variables are subject to deviations due to the actual floor friction and slope as well as the uncertain payload weight. These unknown operation environment parameter influence the vehicle's dynamics and thus affect the controls. These sources for inaccuracies in input- and output-variables cause the interest to apply fuzzy control in this context. Due to the overlap of the different fuzzy classes there result smooth trajectories, as appropriate for payload safety. The discretisation into position classes (fuzzy sets) leads on the other hand to benefits for the processing speed. Another advantage of this approach is the ease in the maintainability of fuzzy controls due to close relationship to linguistic models of the control process, which can be easily interpreted by human operators.

environment and a map of the surrounding. For the robot W ALTER shown in Fig. I eight ultrasonic sensors are used, one directed straight forward, one inclined to +45° and one to -45 0 relative to the main vehicle axis, each two at the left and right side of the robot and one at the back side. While the used Polaroid ultrasonic sensors measure the distance to an object well (± I cm per meter distance), the angular resolution is rather poor (about 20° or 40°, depending on sensor type). Other problem areas related to ultrasonics include • the near range blind zone (at distances below 17 cm) • crosstalk between multiple sensors • from flat surfaces the transmitted beam might be reflected away from the receiver • reaction to hot/cold air stream barriers • absorption at textiles. Thus, ultrasonic sensors provide a low cost solution for distance determination in the range between 20 cm and 10 m, but the limitations recommend them to be supplemented by alternative measurement principles. For our applications mainly an 2 dimensional infrared scanner is considered. Here the resolution is about 0.3 % of the distance, but depend very much on the reflection properties of the object. Limitations of this approach include • the range to the object must be between 40 cm and 2 m • objects must provide a minimum homogeneous reflective area • glass surfaces are hardly detectable • specular or shining objects cause spurious signals (even if outside the measurement range) • it is impossible to detect low obstacles and to avoid disturbing reflections from the floor.

In case unforeseen obstacles are detected on the planned path, the obstacle avoidance algorithm is activated and receives highest priority until the object is passed. Here a compromise between quick reaction capabilities and the suppression of initiating this mode on spurious measurements is to be made. A secondary objective is to perform an efficient detour manoeuvre, without major deviations from the target direction. For this purpose the ultrasonic sensor system, detects objects in a distance up to 3 m and classifies for each sensor into three range classes (close, medium, far) , constituting the fuzzy sets. In order to minimise the reaction time delay during the normal driving operation only the three ultrasonic sensors directed forward are in use for control. The fuzzy input sets for these sensors to the inference

Thus, for control of the robot during collision avoidance a hierarchical strategy is proposed, using a) in a first , fast reacting level the 8 ultrasonic sensors, b) in a second level a rotating ultrasonic sensor (Fig . 3) for verification of the obstacles detected with one of the fixed mounted ultrasonic sensors, but also for classifying of the obstacles, c) in parallel a 2-D-infrared scanner and finally d) a combination of the ultrasonic and infrared sensors, where the measurement data are fused by an occupancy grid method.

..... --~





I

3.1 Fast reacting control level During the normal operation of the robot driving from a starting point to the target a fast reaction capability due to the occurrence of obstacles must be provided . Therefore only the measurement signals of the ultrasonic sensors are used for control. Fuzzy logic provides in this context a robust method to derive from these limited measurements reasonable



y



Fig. 2. Simulation environment for a typical obstacle avoidance situation. In the middle column there are the fuzzy sets of the input - output variables. Above are indicated the steering angle and the speed. 319

Table 2: Typical inference rules for collision avoidance, determining the output variables (velocity and steering angle) to the input variables (range classes of the three sensors and relative orientation). In the diagonal reactions depend in addition on the relative orientation. (Abbreviations : med . = medium, fright = far right. fleft = far left) Sensors: right medium

far

speed is fast angle is right

speed is medium angle is right

close left slow, right slow, left

close

slow, right slow, left

medium

speed is medium angle is left

far

speed is medium angle is left

slow, right slow, left

speed is medium angle is left

speed is medium angle is right med., fright med ., fleft

med., right med ., left

sequentially each of the both sensors transmit ultrasonic waves and receive them again . The measured answers of each sensor are depending on the kind of obstacles. In Fig. 4 shapes of some typical obstacles as they occur in indoor environments, like flat walls or corners or edges, and the corresponding theoretical measurement profiles are depicted .

consist of these range classes and a "relative orientation", defined by the angle between the vehicle's axis and the direction towards the target. Thus the local sensor information for collision avoidance is supplemented by a global path efficiency aspect. There are about 50 rules connecting the fuzzy input sets to the output variables: steering angle and speed of the vehicle . In Table 2 some exemplary rules of the collision avoidance scheme are displayed. As smooth control reactions are desired, for the defuzzification step the centre-of-gravity method was employed . Fig. 2 shows in a simulation environment a typical obstacle avoidance situation using these three sensors.

Due to the characteristics of real sensors the measured profile differs from this form. Especially the total reflection of the ultrasonic waves at an inclination of more than 20°, resulting in the wrong information, that no obstacle is obvious, makes a preprocessing of the measured data necessary. For a corner in Fig. 5 the pre-processed measurement profile of both sensors are depicted. Depending on these profiles and by comparing the shift of both sensor measurements a classification of some typical shapes of obstacles is possible.

Depending on the application also a finer discretisation of the measurement range as well as more sensors (up to 8) can be introduced, but this increases the amount of applicable rules as well as the period required for data processing. Therefore the information of all 8 ultrasonic sensors are only included in the fuzzy strategy if there seems to be an dangerous or doubtful situation indicated by nearby obstacles detected by the three sensors oriented forward .

3.2

slow, right slow, left

The computing time for pre-processing of the measurement signals and a classification of the shape is less than 100 msec on a 80C 166 microcontroller and the time necessary for a scan of 100° is less than 8 sec .

Verification and classification of obstacles by a rotating ultrasonic sensor.

left sensor

right sensor

~-~t¥

If, during this normal operation, obstacles are detected by one of the fixed mounted ultrasonic sensors for verification of this measurement but also for classification of the detected obstacle, a rotating ultrasonic sensor is used. The principle scheme of the sensor is depicted in Fig. 3. Two ultrasonic sensors are mounted in a fixed distance between each other on a beam, which can be rotated by a stepper-motor in steps of some 1°. The rotation can be measured by an angle encoder. The maximum rotation angle of the sensor due to constructive characteristics is some 330°. The sensor has a measurement range of some 20 cm up to 3 m. Depending on the direction of the obstacle detected by the fixed sensors a scan angle is commanded. Within this region in steps of 10

angle

enoode~ ~ 5~-

steppermotor

Fig. 3. 320

Principle of the rotating ultrasonic sensor

a) COlected measuing

:nx> r

''i

I=t 8

,.~

-wi-o

O~.L-_____________ 0

aa et a caner

'0

10

lIC

40

100

w:;. 10

le!

10

11

1fro 1CXXl

:s

fro

c

100 " 0'1'0

-;;

b)

o 11111 111111 111111111111:111111 11 111111111111111111111 11111111111111111111111111111111111111111111 III IIIIIIIIIII'IIIIII!II o v N v v N o ~ N M ~ ~ ~ ~ m ~ ~ N M v ~

~

/'

~

~

~

~

~

~

~

sge(") "1--_ __

~

_ __

Fig. 5.

~

Typical pre-processed signal of the rotating ultrasonic sensors as an answer to a corner ._ _ _ _ Ieft sensor, - - - - - right sensor)

c) e.::nttclllcuw("')

Fig . 4 .

(Elfes, 1989) is implemented . This method integrates different measurements in time and locally by stochastic means . For that the relevant region is discretised in single cells. The measurements of the different distance sensors are evaluated to define the probability that a cell is occupied by an obstacle. The method uses a stochastic sensor model and evaluates the estimation by a Bayes algorithm. The actual measurements of the different sensor types are combined to a resulting occupancy probability of the di fferent cells.

Typical signal of the rotating ultrasonic sensor as an answer to (a) a corner (b) a flat wall and (c) an edge.

The main drawback of this method lies in the high computational effort, which prevents an on-line use during driving. Thus, in practice, if there occurs a doubtful measurement, it can be necessary to stop the robot and wait for the result of the occupancy grid evaluation before the robot can continue his mission . But of course this method is more meaningful than to risk a collision with an obstacle. After this evaluation process the control of the robot can again be continued using fuzzy logic considering the probability measures of the cells. In Fig. 6 a simultaneously representation of the ultrasonic sensors and the 2-dimensional infrared scanner in an occupancy grid is depicted .

3.3 Infrared scanner Due to the limitations of the ultrasonic sensors mentioned above the obstacle detection is supplemented by a 2-dimensional infrared scanner. Especially in complex structured environments and in case of special obstacles there are differences between the measurements resulting from the ultrasonic and the infrared sensors. In extreme cases, for example in front of glass doors, there can be indicated an obstacle by the ultrasonic sensor and no obstacle by the infrared sensor. But there are also situations possible, as for example at flat walls to which the sensor is directed in a large inclination, where the ultrasonic waves are not reflected to the sensor, what means that the sensor doesn't recognise an obstacle. But the infrared sensor would indicate a possible collision .

4 CONCLUSIONS

The main topic of this paper was the development of simple procedures for control of a mobile robot for obstacle avoidance using low-cost sensors. The loss in accuracy by using these sensors has been partly compensated on the one side by increasing the numbers of sensors and on the other side by implementing different types of sensors and fusion of the measurement data . It has been proven that Fuzzy Logic is a useful method for robot control as well as for simple data fusion.

Thus, especially for such doubtful cases, it is necessary to fuse the measurement data of the different sensors in order to increase the information contents and to provide a more sure control of the robot. For these cases a method based on occupancy grids

321

-- I

l : ::: :: ~ -

.Ij :

Fig. 6. Simultaneous representation of the ultrasonic sensors (lower left part of the figure) and the 2 dimensional infrared scanner (upper left part) in an occupancy grid (right part) .

For more complex environments more sophisticated sensor systems are necessary and also the information processing methods have to be adapted. In this paper on a low level the measurement data of different ultrasonic and infrared sensors have been fused by an occupancy grid method using statistically based tools and on the upper level of the control hierarchy the results have been evaluated using fuzzy logic.

REFERENCES Abidi, Mongi A. / Gonzalez, Rafael C. (eds.) (1992). Data Fusion in Robotics and Machine Intelligence, Academic Press Doyle,R. and H. Roth (1995). Das Kalman Filter in: Handbuch der Sensonechnik. Obermaier, E. / Trankler, H.-R. (eds.). Springer VerJag. Elfes, A. (211989). Using Occupancy Grids for Mobile Robot Perception and Navigation. IEEE Computer Magazine. Kokar, M. , K. Kim (eds.) (211994). Special Section on "Data Fusion", Control Engineering Practice . 803-897 . Schilling, K. (1995). Signalverarbeitung bei Multisensoren. in: Handbuch der Sensonechnik. Obermaier, E. / Trankler, H.-R. (eds.). Springer VerJag.

ACKNOWLEDGEMENTS The project has been supported by the German Federal Ministery for Education and Science and the Ministery for Science and Research in BadenWiirttemberg, Germany.

322