TOWARDS CONCEPT FORMATION GROUNDED ON

PERCEPTION AND ACTION OF A MOBILE ROBOT

Volker Klingspor and Katharina Morik

University of Dortmund, Computer Science Dept. LS VIII, D-44221 Dortmund, Germany

fvolker,[email protected]

Abstract. The recognition of objects and, hence, their descriptions must be grounded in the
environment in terms of sensor data. We argue, why the concepts, used to classify perceived objects
and used to perform actions on these objects, should integrate action-oriented perceptual features
and perception-oriented action features. We present a grounded symbolic representation for these
concepts. Moreover, the concepts should be learned. We show a logic-oriented approach to learning
grounded concepts.
Key Words. Operational concepts, inductive logic programming, combining sensing and action,
symbol grounding

1. INTRODUCTION oriented towards

levels of perception action

Up to now, application programming of a robot is

control

a time consuming task. The programming is car-

concept structure
planning
ried out at a very low level. The goal position of
the individual commands must be given by exact navigation operational concepts

real world coordinates. This has three undesired

consequences. First, even if the new application is path, motion
sensor control
reflexes

very similar to the one before, the control program sensors effectors

must be rewritten completely. Second, these com-

mands require a xed environment and complete
knowledge of the environment. If either an un- Fig. 1 Levels of control
foreseen event occurs or the environment is (par-
tially) unknown, this approach to preparing an
application fails. Third, communication between Machine learning can be applied to robot tech-
a knowledgeable user and a robot is hindered by nology in several ways. Consider a hierarchical
the low-level representation of the application pro- control structure with dierent levels of behaviors
gram. The user cannot change the robot's behav- as it is proposed by (Brooks, 1986) (Fig. 1). At
ior directly according to small changes in the envi- all levels, perception-oriented and action-oriented
ronment, nor can he help the robot out of a failure processing is related. Machine learning can en-
situation. hance the performance at each level.
One way to overcome the problems described Biologically inspired work in robotics has devel-
above is to develop robots that learn. From its ex- oped articial beings that adapt to their environ-
perience in an environment the robot could adapt ment (Brooks, 1991; Steels, 1993). This type of
its behavior to the environment. Moreover, the learning is restricted to re ex-like behavior. Simi-
robot could transfer what it has learned in one larly, the system described by (Millan and Torras,
environment to other, similar environments. A 1992) operates on the re exes level. It learns how
second way is to use less specic commands. If, to avoid obstacles and reacts very fast to sensory
for example, the user wants a mobile robot to leave input, but higher levels of cognition such as rea-
the room, turn to the right until it reaches a desk, soning and concept formation are excluded. The
and then stop, it should be possible for the user to system is unable to nd a path in a maze-like en-
state just this. This requires the robot to be ca- vironment in an appropriate time. Nevertheless,
pable of recognizing doors and { in the absence of learning at this lower level enables the system to
specic goal coordinates { be capable of searching react immediately to unexpected situations and
for them. The robot must assign its perceptions thus decreases eort at higher levels.
to classes of objects, i.e., to concepts (e.g., doors).
Our aim is to combine both of the described ap- In order to enhance a robot's high-level pro-
proaches to build up more applicable and more cessing, planning can be tailored to particular
exible robots in future. robot tasks by learning techniques (Segre, 1988).
Learning from failures of robot actions can be object, this object is an instance of the concept,
used to enhance the planning capabilities of a otherwise it is not an instance.
robot (Zercher, 1992; Bennett, 1989). Learning
can also be used to link the environment model Our topic is to learn concept descriptions for mo-
with the perceptions when executing a plan (Jong bile robots. The input data used for learning are
and Bennett, 1993; Gil, 1994), thus adapting a sensor measurements. In contrast to most prob-
general plan to the specic environment. lems examined in symbolic learning, no explicitly
given features exist, only real values are given.
Our approach to learning robots links low-level But this data is too specic to be applied to more
features that integrate sensing and action with than a single object. Thus, more abstract features
high-level concepts. These concepts, which we must be constructed from the sensory data. The
call operational concepts combine sensing and ac- choice of these features and their calculation is
tion and are grounded on basic features. The use a very important task, because it in uences the
of these concepts sets up the following learning expressiveness of the symbol system. Concepts
tasks: that cannot be described by these features can
neither be described by newly built intermediate
 learning to recognize perceived objects as in- concepts, since intermediate concepts only sim-
stances of a concept, plify the representation, but do not enlarge the
\closure of existing symbols" (Wrobel, 1991).
 learning to enhance a map of the environment
by adding instances of perceived concepts, Until now, feature construction was programmed
 and learning which sensing and moving ac- by an application developer. After this program-
tions are appropriate for easing an object's ming step, he had to test the quality of the
recognition. constructed features with the used system which
might be a machine learning system, and then he
The learned concepts can then be used to formu- had to change the way features are constructed
late high-level commands and to control their ex- if the performance of the system were not su-
ecution (Wallner et al., 1994). cient. This process is very time consuming, and
if the environment or the robot changes, feature
In this paper, we focus on two main properties of construction must be adapted to this change.
operational concepts: how they can be grounded Feature construction cannot be seen as an isolated
in the environment (Sect. 2.1) and how to inte-
grate perception and action (Sect. 2.2). In Sec- task. If the concept learning step is not taken into
tion 3, we describe the data used for the experi- account while developing the feature construction,
ments. Section 4 focuses on the representation of the feature construction step cannot get any ex-
the data and the applied learning techniques. Fi- ternal evaluation about the quality of the con-
nally, we discuss open problems and restrictions structed features. So learning how to construct
of our approach (Sect. 5). features must be combined with concept learning.
Additionally, the features must be delivered on-
2. REQUIREMENTS OF OPERATIONAL line, because the robot must be able to react im-
CONCEPTS mediately to its perceptions. Thus, the algorithm
that calculates these features must be incremen-
Two requirements were to be regarded during the tal, analyzing each new input as soon as it is given.
development of representing operational concepts.
First, the concepts should be usable by real robots For learning the basic features, a set of functions
in a real world. Therefore, the features used to de- and their possible parameters are given. The func-
scribe the concepts must be grounded in the en- tions calculate basic features on the basis of sensor
vironment (Harnad, 1990; Wrobel, 1991; Cottrell measurements. The system learns the appropriate
et al., 1990). Second, inspired by psychological degree of granularity by searching for an appropri-
results, we argue why the representation should ate function and a useful set of parameters. The
integrate perception and action. search space is partially ordered by the granularity
of the functions and parameters. The granularity
is measured in terms of how many feature state-
2.1 Symbol Grounding ments are delivered for a series of measurements.
If the measurements are described by few feature
From the AI point of view, concept characteriza- statements, it is a rough, highly compressing ab-
tion means the denition of a concept based on straction. If many feature statements are deliv-
features, so that examples for this concept will ered for the same series of measurements, it is a
be covered by the characterization. If a formula nely grained, less compressing abstraction. The
representing this characterization, is true for an
appropriateness of the granularity is measured in The main issue of a cup is that you can drink from
terms of the quality of learning results using the it. If a cup is dened as a receptacle from which
features. If no good concept learning was possi- drinking must be possible, the object classica-
ble using a rough abstraction, a ner granularity tion can be veried by performing this action. If
is tried, and so forth, until a good learning re- it is possible to drink from the perceived object,
sult has been achieved or no further function and it can be classied as a cup. In this way, actions
parameter set is available. are integrated into concept descriptions and their
recognition functions.
2.2 Integration of Perception and Action Kedar-Cabelli has gone a rst step in this direc-
tion (Kedar-Cabelli, 1988), using properties that
The second main issue of operational concepts is describe action applicability, e.g., that a cup must
the integration of perception and action, where be liftable. But the applicability is described by
perceptual features must be action-oriented and perceptual features such as material, weight, han-
action features must be perception-oriented. Hu- dle, etc., { not by action features proper. Gior-
mans represent events by relating the objects in- dana and Saitta have proposed to use executable
volved and the action, that is performed at these features in concepts descriptions (Giordana and
objects. Abelson called predicates what comprises Saitta, 1990). These features are true for an ob-
an action and its corresponding object (Abelson, ject, if a particular handling of this object is suc-
1963), e.g., drinking from a cup, throwing a ball, cessful. For instance, the feature \movable" for a
moving through a doorway. Nelson states, that concept can be veried by moving that object. We
\ , young children must represent their own want to develop their approach further and pro-
:::

roles and the roles of other and be able to recip- pose that features at all levels should integrate ac-
rocate actions of the other with actions of their tion and perception. Perceptual features require
own, " (Nelson, 1983, p. 135). Objects should the integration of the action that is performed
:::

be represented as relations between actions and while perceiving an object. Suppose that you are
reactions, where an object implies a specic ac- looking at a cup from above. This position does
tion, like a \ball implies throwing". Nelson re- not allow to determine whether the bottom is at
gards three kinds of relations between objects and or not. The perception is restricted by the action
actors, where the two higher levels are the most during which an object is perceived. Action fea-
appropriate ones in our scenario. In the medium tures, in turn, require the integration of percep-
level, \dierent objects may occur in the same po- tion. Particular sensor patterns express the appli-
sition of an event", like a specic tennis ball or a cability conditions, the conditions for successful
red rubber ball in the event \ball implies throw- performance of the action, and the conditions for
ing". In our representation, the dierent objects ending the action. In this way, an action is ex-
that can occur in an event are represented by their pressed by perceptions.
perceptual features, and the action these objects
imply are integrated in the denition of opera- 3. DATA FROM THE NAVIGATION
tional concepts as action features. Additionally, SCENARIO
\an object or category of objects may recur in a
number of dierent events, in similar or dierent Before describing the representation hierarchy and
relationships". The category of doors may occur the learning algorithms, we present the scenario
in dierent events like moving along them or mov- and the robot we used for getting data. PRI-
ing through them, so we have to relate dierent AMOS, developed by the University of Karl-
actions with the perceptions of the concept door. sruhe (Dillmann et al., 1993), is a mobile robot
with three degrees of freedom for motion, i.e., it
This integration of perceptual features and action can move into every direction and rotate simul-
features also supports the classication of an ob- taneously. It has 24 sonar sensors measuring the
ject as an instance of a concept. In a conventional distance to the nearest object within the emission
representation, a cup, for example, is dened by cone. Errors can occur, if the sonar beam is mul-
having a at bottom and a handle and being con- tiply re ected or the angle between the beam and
cave. But it is easy to nd arbitrarily many ob- the object is inconvenient. At every side of the
jects with these properties that are not cups be- robot three sensor are installed, emitting in par-
cause it is impossible to drink from them, e.g., if allel. Three sensors are mounted at every corner
the handle bridges over the opening (De Jong and emitting with a dierence of 15 .
Mooney, 1986). Finding a complete description
of a cup excluding all exceptions is impossible be- The aim of our rst learning trials was to learn
cause of the innite number of these exceptions. descriptions for the concepts move along door and
This is the qualication problem (McCarthy and move through door. We used data from 28 traces
Hayes, 1969). So, how to dene a cup suitably? in a simple room, most of them being paths
 and their relations. The classical learning algo-
rithms for attribute-value representations cannot
solve our learning tasks.
10 2
9 3

4.1 Learning Basic Features

8 2 4
1 3
0

In Section 2.1 we have argued that the appropriate

construction of basic features is crucial for further
concept learning. Moreover, we have explained
the necessary interaction of feature construction
11 7 5 1

and concept learning. Now we present the learn-

ing algorithm for feature construction, starting
with the specication of the learning task:
6
Learning basic features:
0
Given a set of functions and parameter
Fig. 2 Room with traces and edge numbers values of these functions and an ordering
on these dierent ways to calculate basic
features {
along or through a doorway in dierent directions Try to nd the function and parame-
(Fig. 2). The cupboard, displayed as the hatched ters that lead to good results of concept
object, appeared only in some of the traces. For learning.
every trace, we got between 25 and 35 measure-
ments of the 24 sensors, summing up to 17472 in Our learning task is to select appropriate values
total. Each measurement consists of the following from a given set. This is simpler than the con-
data: struction of dierent functions for calculating ba-
sic features. The goal of learning is to nd the pa-
 trace number and point of time; for identi- rameter setting, whose application results in ba-
cation of the measurement, sic features with the appropriate level of detail.
 robot position in world coordinates, In this context \appropriate" means, that, on the
 sensor number, its position and orientation; one hand, the basic features contain only those
for identication of the sensor, details, which are necessary to recognize and dis-
tinguish operational concepts. On the other hand
 measured distance, it means, that they do not contain super uous,
 position of the sensed point in world coordi- redundant details, which would impair the recog-
nates, nition process by overloading it. The set of pa-
 object and edge number of what is sensed. rameter settings can be considered as a version
space (Mitchell, 1982) for learning basic features.
The learning algorithm can be described by the
following steps:
4. THE REPRESENTATION AND
LEARNING APPROACHES
1. Start with a predened set of parameters.
Learning can take place at several levels within 2. Calculate the basic features based on the
a concept hierarchy representing concepts at dif- function determined by the actual parame-
ferent degrees of abstraction. In this section, we ters.
rst present our approach to overcome the sym-
bol grounding problem, namely the learning of 3. Learn concept descriptions.
basic features (Section 4.1). Then, we give an 4. If the evaluation of concept learning is not
overview of the representation hierarchy, based on acceptable and other parameter values are
the basic features. Finally, we show how learn- available, take the next, less abstract param-
ing can be applied to higher levels of the hier- eter setting and iterate at step 2.
archy. The used learning algorithm will not be Otherwise: stop.
described in detail (but see (Klingspor, 1994)).
The reader may think of any learning algorithm
for (restricted) rst-order logic because, in prin- The learned (i.e. selected) functions and parame-
ciple, any rst-order learning algorithm can solve ters are used to calculate basic features from sen-
our learning tasks. Note, that rst-order learn- sory input. The calculation is incremental so that
ing is necessary in order to handle time intervals it can be performed while the robot is moving.
 operational concepts
Basic perceptual features:
increasing(trace1, sensor1, t1, t4, 0.7). perception−integrating actions
incr_peak (trace1, sensor1, t4, t5, 3.2). applied to learned from

Measured Distance
decreasing(trace1, sensor1, t5, t7,−0.3).
no_measurement(trace1, sensor1, t7, t10, _).
stable (trace1, sensor1, t10, t13, 0.01). perceptual features
3m
sensor features action features
basic features
2m

sensing and action

1m

t1 t4 t5 t7 t10 t13 Fig. 4 Representation Hierarchy

Time

Fig. 3 Measurements and basic features. maximal sensing distance. This situation, in the
example the interval from t7 to t10, is labeled
no measurement. Other basic features are stable
Basic features are calculated for each sensor using (describing more or less equal measurements), de-
the following algorithm: creasing, decreasing peak, no movement (we can-
not compute the gradient, if the robot does not
1. Start with the rst measurement. move), and straight to and straight away (if the
2. Get the next measurement and calculate the gradient is about 1 or ,1, respectively).
gradient between the previous and the cur-
rent one, i.e., the quotient of the dierence of Note, that the basic features compress a sequence
the measurements and the moved distance. of measurements to a statement about a time in-
1 and 2 becomes an interval.
terval. The nely grained basic features summa-
rize only a short sequence of measurements where
T T

3. If the gradient is close enough to the average the rough features summarize a long sequence of
gradient of the interval: measurements. In that way, the basic features de-
then enlarge the interval by the new time termine the time intervals used by concept learn-
point and adapt the average gradient de- ing. The time intervals, in turn, determine what
pending on the way dened by a param- can be achieved by concept learning. It is, for in-
eter. stance, impossible to learn good concepts descrip-
else close the previous interval at the previ-
tions, if the time intervals determined by the basic
ous time point and start a new one at features do not match the time intervals used in
that time point with the new gradient the examples to be classied.
as average gradient.
\Close enough" is dened by a parameter, 4.2 The Representation Hierarchy
too.
4. A closed interval will be named by a symbol, Several levels of abstraction link the level of raw
depending on the average gradient attached sensor and motion data with the level of opera-
to that interval. The set of symbols and the tional concepts. The hierarchy of more and more
gradients they belong to are determined by a abstract representations comprises perceptual fea-
further parameter. tures at several levels which are all related with
5. Go to step 2. actions as well as action feature which become in-
tegrated with perceptual features at a higher level.
Figure 4 shows the representation hierarchy.
In Figure 3, an example for a short sequence of
measurements is displayed together with the cor- In the rst step, basic features are constructed
responding basic features. First, the gradient be- based on the measurements and the robot's po-
tween two measurements is more or less 0 7. The : sition. At the next higher level, the perceptions
measurements are increasing linearly. That is while moving are represented by single sensor fea-
represented by the basic feature increasing(trace1, tures.
sensor1, t1, t4, 0 7). Then, from time point t4
:

to t5, the gradient is greater than 1. Since the s line: a single edge.
dierence between two measurements cannot be s jump: two parallel edges, sensed in sequence.
greater than the moved distance, if the sensor
sensed the same edge, two dierent edges must s convex: two orthogonal edges, sensed from the
have been sensed. This situation is expressed by convex side of this corner.
the basic feature incr peak. Sometimes a sensor s concave: two orthogonal edges, sensed from the
gets no echo, e.g., if there is no object within the concave side.
Each sensor feature is characterized by sequences verication conditions, and end-of-action condi-
of basic features. The sequences are represented tions for an abstract action and its corresponding
by rules. An example rule for s jump is: object.
stable(Trace, Or, Sensor, Time1 , Time2 , Grad1 ) & standing(Trace, T1, T2, in front of door,
incr peak(Trace, Or, Sensor, Time2 , Time3 , Grad2 ) & PDirect, small side, PrevP) &
stable(Trace, Or, Sensor, Time3 , Time4 , Grad3 ) parallel moving(Trace, T2, T3, MSpeed,
! s jump(Trace, Sensor, Time1, Time4, parallel). PDirect,through door,right and left) &
standing(Trace, T3, T4, in front of door, rear,
At the next level, situations are represented where small side, through door)
dierent sensors of a class sensed the same con- ! move through door(Trace, T1, T4).
stellation of edges while moving. Single sensor fea-
tures are grouped together in order to increase the The rule describes the operational concept of mov-
evidence for an edge constellation. This helps to ing through a doorway. It combines the recogni-
cope with the noise of single sensor measurements. tion of the doorway with the action of moving
It can be sucient for a sensor group feature to be through it on a parallel track. Moving diagonally
derived, that two of three sensors of a sensor class through a doorway is a dierent operational con-
sensed the same constellation. Additionally, we cept. Action and perception are linked by the
abstract from the actual sensor numbers in this perceptual features occurring as arguments in the
step. Instead, the orientation of the perception action predicates. In our example rule, the per-
relative to the orientation of the motion is used. ceptual features are written in italics.
s jump(Trace, Sensor1 , Start1 , End1 , parallel) & The rst premise of the rule states that the robot
sclass(Trace, Sensor1 , right side) & is standing in a time interval from T1 to T2 of a
s jump(Trace, Sensor2 , Start2 , End2 , parallel) & particular (Trace) and senses the perceptual fea-
sclass(Trace, Sensor2 , right side) & ture in front of door from the sensors of a small
succ(Start1 , Start2 ) & succ(End1, End2 )
! sg jump(Trace, right side, Start1 , End2, parallel). side after having perceived the perceptual feature
PrevP prior to T1. This premise denotes the pre-
At the next level of abstraction, the level of action- condition for moving through a doorway on a par-
oriented perceptual features, perceptions at dier- allel track. The precondition can also be viewed
ent sides of the robot will be combined. as a trigger for planning: whenever recognizing
that it stands in front of a door, the robot may
execute the action of parallel moving through the
sg jump(Trace, right side, T1 , T2 , parallel) & doorway.
sg jump(Trace, left side, T1 , T2 , parallel)
! through door(Trace, T1, T2, both sides, parallel). The second premise states that in a following time
interval the action parallel move is executed, mea-
At the action side of the representation hierarchy, suring by the sensors at the right and the left side
only basic actions occur. They represent intervals of the robot the perceptual feature through door.
of uniform motions. From the absolute orientation Note, that the particular values for the time in-
and speed of the motion features are calculated tervals do not matter for the operational concept
that relate the orientation to the previous direc- but are instantiated by the specics of a particular
tion of motion. The basic action features used are path. Only the time relation is important. This
move, rotate, and stand, with the arguments trace, makes the operational concept independent of the
time interval, speed, and direction. particular depth of a doorway and the robot's par-
ticular rate of advance.
Action features and perceptual features are then
combined to perception integrating features repre- The third premise describes the end of the move-
senting single actions and the corresponding per- ment through the doorway: the robot is stand-
ceptions. An example of such an action is a steady ing in front of the door, now sensing with its
movement while measuring through door in a spe- rear sensors mounted at its small side the feature
cic direction PDirect. in front of door after having perceived the feature
through door. PDirect, the orientation of percep-
through door(Trace, PDirect, T1 , T2 , parallel) & tion (e.g., left, right, front, rear) is not xed by
move(Trace, T1 , T2 , Speed, Dir) the rule. It is only relevant that the orientation
! move parallel(Trace, T1 , T2 , Speed, Dir, should not change during the parallel move and
through door, PDirect). that with respect to this orientation the doorway
is sensed in the rear after the movement. That
Finally, operational concepts use perception inte- means, whatever the orientation was at the be-
grating features for characterizing preconditions, ginning of the action, it is dened as being front
so that after completing the action the doorway is creases as the representation hierarchy is climbed.
sensed by the opposite, the rear sensors. This shows that the hierarchy makes learning ro-
bust against the noise which is produced by the
weak ultrasonic sensors. The levels can be re-
4.3 Learning Perceptual Features garded as a lter { they only pass the more reliable
information to the next higher level.
Learning of higher-level features and, nally,
of operational concepts is performed in the
paradigm of inductive logic programming (ILP, 5. CONCLUSION
see, e.g., (Muggleton, 1992)). The learning task is
the following: In this paper, we introduced operational concepts
that integrate action-oriented perceptual features
and perceptual-oriented action features. Both
Learning higher-level features:
the target feature for various sit- psychological and practical arguments indicate
Given
uations and background knowledge { the necessity of this integration. We sketched
learn rules that characterize the tar-
a representation for operational concepts that is
get feature in terms of the background grounded in the perception and action of a mo-
knowledge. bile robot. We proposed two learning tasks: one
that nds appropriate basic features in order to
summarize sequences of sensor data, and one that
The background knowledge is a set of facts from characterizes higher-level features (or operational
the level below the target feature and general in- concepts) in terms of features from the level below.
formation about the robot (e.g., sensor classes). Experiments have shown that learning becomes
This learning task is performed several times: more accurate and the coverage of learned rules
learning sensor features from basic features, learn- increases when progressing to higher levels of the
ing sensor group features from sensor features, representation hierarchy. This means that even
learning action-oriented perceptual features from noisy input to learning leads to results that sum-
sensor group features, and learning operational marize the data such that at the next higher level
concepts from perception-integrating action fea- the noise is reduced. These results are promising,
tures. but further learning experiments in other environ-
ments need be conducted.
We applied learning algorithms that search a re-
stricted hypothesis space completely. The hy- In order to achieve rst results, we restricted sens-
pothesis space is restricted syntactically by rule ing and action of the mobile robot. For rep-
schemata that express the form of learnable rules. resenting the sensory data, we reduced the de-
Rules that are not instantiations of a rule schema grees of freedom of the robot allowing only one-
cannot be learned. This prohibits learning rules dimensional motions at once. That will lead to
that cannot be put to use by the robot. The rule problems, if we incorporate a reactive module ex-
schemata are ordered according to their general- ecuting the elementary operations, if this module
ity. Starting with the most general rule schema, generates traces where the robot deviates from
all its instantiations (i.e. rules) are tested whether straight courses. Then, the calculation of basic
they cover positive but not negative examples of features is probably not able to counterbalance
the target feature. Until no negative (or only few) the divergence of the sensor measurements, be-
negative examples are covered, the next special cause in the moment, we only use linear functions
rule schema is instantiated and tested. The most to calculate the features.
applicable learning algorithms were RDT (Ki-
etz and Wrobel, 1992) and a modication of it, Another restriction concerns the sensors used.
GRDT (Klingspor, 1994). Sonar sensors are not able to get precise mea-
surements, they cannot be used to discriminate
In a nutshell, the results are as follows. Given ne structured objects. E.g., it is not possible
1004 examples for the four sensor features, we to detect closed doors, so we used only simple
learned 129 rules, covering 87% of the given exam- structured objects in the moment. An interest-
ples. For learning sensor group features, we had ing future investigation might be the integration
given 956 examples, from which GRDT learned of another sensor system like a vision system to
136 rules. Using the learned rules for sensor fea- get further information in specic situations.
tures and sensor group features, we got a coverage
of 64%. For the operational concept through door, The next challenge will be to validate the con-
we had given ten examples. We learned three tribution of learning to easing robot applications.
rules, two of them covering exactly the ten ex- Until now we have a small planning component
amples. The third rule was a poor one that could that computes basic actions of a robot from an
be dropped. The quality of the learned rules in- operational concept. This component uses learned
rules in a backward-chaining manner. The basic Giordana, A. and Saitta, L. (1990). Abstraction {
perceptual features are incrementally calculated a general framework for learning. In Procs. AAAI
from sensor measurements as shown in this pa- { Workshop on Automatic Generation of Approxima-
per (Section 4.1). Using the learned rules in a tions and Abstractions.
forward-chaining manner, higher-level features are Harnad, S. (1990). The symbol grounding problem.
derived. Using these techniques, learned rules are Physica D, 42:335{346.
made available for the robot's processing. First, Jong, G. D. and Bennett, S. (1993). Permissive plan-
experiments with the robot are to be made where ning { a machine learning approach to linking internal
the robot uses learned rules in the environment in and external worlds. In AAAI-93.
which learning took place. Then, the same rules Kedar-Cabelli, S. (1988). Toward a computational
should be used by the robot in similar but dier- model of purpose-directed analogy. In Prieditis, A.,
ent environments from the one in which learning editor, Analogica. Morgan Kaufmann.
took place. Finally, the eorts of preparing these Kietz, J.-U. and Wrobel, S. (1992). Controlling the
robot applications need be compared. complexity of learning in logic through syntactic and
task-oriented models. In Muggleton, editor, Inductive
ACKNOWLEDGEMENTS Logic Programming. Academic Press.
Klingspor, V. (1994). GRDT: Enhancing model-based
This work is partially funded by the European learning for its application in robot navigation. In
Community under the project B-Learn II (P7274) Wrobel, S., editor, Proc. of the Fourth International
and the Ministry for Sciences and Research of the Workshop on Inductive Logic Programming, GMD-
German federal state Nordrhein-Westfalen. Studien Nr. 237, St. Augustin, Germany.
We would like to thank Anke Rieger for her con- McCarthy, J. and Hayes, P. J. (1969). Some philo-
sophical problems from the standpoint of articial in-
tributions to representing and learning perceptual telligence. Machine Intelligence, 5:463{502.
features, Stefan Sklorz for developing the repre-
sentation for action features and programming a Millan, J. and Torras, C. (1992). A reinforcement
meta-interpreter for planning, and Stephanie Wes- connectionist approach to robot path nding in non-
maze-like environments. Machine Learning, 8.
sel for her work on learning basic features. Thanks
to Joachim Hertzberg and Stephanie Wessel for Mitchell, T. M. (1982). Generalization as search. Ar-
reading drafts of this paper. ticial Intelligence, 18(2):203{226.
Muggleton, S. (1992). Inductive Logic Programming.
Academic Press, London.
REFERENCES Nelson, K. (1983). The derivation of concepts and cat-
Abelson, R. P. (1963). Hot cognition. Computer Sim- egories from event representations. In Scholnick, edi-
ulation of Personality. tor, New Trends in Conceptual Representation: Chal-
Bennett, S. W. (1989). Learning uncertainty toler- lenges to Piaget's Theory? Lawrence Erlbaum.
ant plans through approximation in complex domains. Segre, A. (1988). Machine Learning of Robot Assembly
Tech. Rep. UILU-ENG-89-2204, University of Illinois Plans. Kluwer, Boston.
at Urbana-Champaign. Steels, L. (1993). Building agents out of autonomous
Brooks, R. A. (1986). A robust layered control system behavior systems. In Steels and Brooks, editors, The
for a mobile robot. IEEE Journal of Robotics and 'articial life' route to 'articial intelligence' { Build-
Automation, 2. ing situated embodied agents. Lawrence Erlbaum.
Brooks, R. A. (1991). The role of learning in au- Wallner, F., Kaiser, M., Friedrich, H., and Dillmann,
tonomous robots. In Valiant and Warmuth, editors, R. (1994). Integration of topological and geometrical
COLT'91, Proc. of the fourth Annual Workshop. Mor- planning in a learning mobile robot. In Proc. of IROS-
gan Kaufmann. 94.
Cottrell, G. W., Bartell, B., and Haupt, C. (1990). Wrobel, S. (1991). Towards a model of grounded con-
Grounding meaning in perception. In Marburger, cept formation. In Proc. 12th Int. Joint Conf. on AI.
editor, GWAI{90, 14th German Workshop on AI. Morgan Kaufman.
Springer.
Zercher, K. (1992). Wissensintensives Lernen fur
De Jong, G. and Mooney, R. (1986). Explanation- zeitkritische technische Diagnoseaufgaben. inx, Sankt
based-learning: An alternative view. Machine Learn- Augustin.
ing, 2(1):145{176.
Dillmann, R., Kreuziger, J., and Wallner, F. (1993).
PRIAMOS { an experimental platform for re exive
navigation. In Groen, Hirose, and Thorpe, editors,
IAS-3: Intelligent Autonomous Systems. IOS Press.
Gil, Y. (1994). Learning by experimentation { incre-
mental renement of incomplete planning. In Cohen,
W. and Hirsh, H., editors, Procs. 11th Int. Machine
Learning Conference.