Online Inference of Robot Navigation Parameters from a Semantic Map

Benjamin Kisliuk

, Christoph Tieben

, Nils Niemann

Christopher Br

ocker

, Kai Lingemann

and Joachim Hertzberg

1,2

German Research Center for Artiﬁcial Intelligence, Plan-based Robot Control Group, Osnabr

uck, Germany

Universit

at Osnabr

uck, Knowledge Based Systems Group, Osnabr

uck, Germany

Keywords:

Semantic Map, Navigation, Robot Autonomy, Semantic Navigation.

Abstract:

Agriculture is becoming one of the key application ﬁelds for mobile robots. At the same time it poses se-

rious challenges for true autonomous systems due to its heterogeneous and dynamic nature. To act robustly

and reliably, robotic behaviour needs to be controlled by an intelligence, making explainable and informed

decisions based on knowledge of its surroundings. However, this knowledge cannot only be derived from sen-

sor data but has to be based on prior knowledge and external sources as well to comprehensively represent a

robots deployment site. By representing this knowledge in formal and thus machine readable way, automated

inference improves the handling of the complex nature of these requirements. In this paper, we show how

quantitative and qualitative control parameters regarding a mobile robots navigation can be derived from a

manually modelled semantic map of an agricultural deployment site. Also we describe how such a system

can be integrated into a typical ROS system architecture. By making the derived knowledge easily available,

the robotic system is enabled to dynamically adapt route planning on an agricultural deployment site and to

switch between different local planning algorithms according to situational and prior knowledge.

1 INTRODUCTION

In agriculture, robots are widely considered an impor-

tant building block to the future of a more sustainable

source of food and primary production with various

projects demonstrating success, mostly in singular ap-

plications (Bergerman et al., 2016).

Automation can help streamlining processes, and

autonomous machines can enable processes which are

too complex to be managed efﬁciently by humans.

Current research projects often focus on multi-robot

systems, which act and decide partially or completely

without human supervision (Shamshiri et al., 2018).

As of today, most commercially available robots

are able to execute certain agricultural processes, but

still lack the capability of an integrated approach

(Bergerman et al., 2016; Shamshiri et al., 2018). This

can be partially attributed to the complexity of the do-

main and the multitude of conditions and processes

required to run an agricultural business; but also the

agricultural sites, i.e. farms, are usually very hetero-

geneous spaces, which are – from a roboticist’s point

of view – far from well structured in a robotics sense

and dynamic by nature (Kunze et al., 2018; Egerstedt

et al., 2018). Thus, a robot which is supposed to run

autonomously will be challenged to deal ﬂexibly with

a lot of problems and steep constraints regarding reli-

ability and robustness while behaving and deciding in

a comprehensible and explainable way (Langley et al.,

2017).

In practice, integrating robotic software often

means that it is up to the roboticist to decide which

approach or existing implementation of a certain tech-

nology – like navigation or task-planning and their re-

spective parameters – ﬁt the overall constraints best.

A ﬂexible approach allowing the robot itself to au-

tomatically choose between different approaches ac-

cording to which ﬁts best in a speciﬁc situation can

improve the ability of a robot to interact with a com-

plex environment enormously like the Move Base

Flex framework for navigation (Putz et al., 2018).

From that ﬂexibility however, the question arises how

to choose the best algorithm or control parameters for

a given situation or context.

This paper demonstrates the usage of a reasoning

engine over a semantic map of the geospatial envi-

ronment as a key component to tackle challenges of

autonomous robot control. We will show how seman-

tic knowledge about the area the robot is deployed in

– the deployment area – is modelled by annotating

156

Kisliuk, B., Tieben, C., Niemann, N., Bröcker, C., Lingemann, K. and Hertzberg, J.

Online Inference of Robot Navigation Parameters from a Semantic Map.

DOI: 10.5220/0010790200003116

In Proceedings of the 14th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2022) - Volume 1, pages 156-163

ISBN: 978-989-758-547-0; ISSN: 2184-433X

a geospatial map with semantic knowledge. The re-

sulting semantic map in combination with informa-

tion sources like weather data, the current time of day

and the robots position can be used to infer control pa-

rameters online. For representation we use a custom

software called SEMPR with an integrated rule based

reasoner using the RETE pattern-matching algorithm

(Forgy, 1982).

We will show as an example how 2D navigation

cost maps can be generated from the inferences and

used for path planning in a robotic system.

After this introduction to the problem the second

chapter will provide information about the theoretical

background and state of the art. In the third chap-

ter we describe the implementation of the necessary

components and show how they integrate into a ROS

robot architecture in chapter four. Lastly, we summa-

rize the system and show prospective future avenues.

2 BACKGROUND

In modern times mobile robot control architectures

commonly base a robots behaviour on the input of

the sensors and the results of some deliberation pro-

cess. However, not every relevant information can or

will be sensed by the robot: In the agriculture domain

properties like ground moisture, dust or the quality

of daylight often will not be sensed due to the lack

of speciﬁc sensors or appropriate software modules.

Some types of information cannot be derived from

on-board sensory data at all, e.g. the legal situation

regarding ownership of a plot or a weather forecast.

This motivates the use of an environment representa-

tion which covers information about the deployment

site which can be enriched with prior knowledge. To

further enhance the autonomy of the robot control

software, this knowledge should be expressed in a for-

mal way which allows it to be interpreted by a reason-

ing software.

In (Hoellmann et al., 2020) a simple approach to

handle different contexts which cannot or will not

be sensed by the robot is described: By dividing an

overall heterogeneous deployment site into subareas

with in themselves higher homogeneity like single

ﬁelds, yard areas, buildings or pasture areas, zones

with more speciﬁc constraints were deﬁned. Based on

the context deﬁned by a zone, control parameters and

navigation algorithms were chosen according to pre-

deﬁned parameters. However, the limitations of this

approach became clear quite early. While the context

based approach was able to tackle the heterogeneity

of the deployment site, it is not ﬂexible enough to re-

gard dynamic changes or complex interconnections.

A more comprehensive approach can be taken by pre-

deﬁning only singular facts, making them changeable

at runtime and using a rule based reasoning system

to infer the control parameters for the robot at run

time. This can be understood as three main compo-

nents: The semantic map which incorporates geospa-

tial anchored knowledge, a set of formalised rules and

a reasoning algorithm.

From intuition, this approach offers multiple ad-

vantages: Independent knowledge can be modelled

independently. This allows for discussing relevant in-

formation about a deployment area with experts in the

domain using non-technical terms in the interviews.

Also the complexity of the system architecture can be

reduced drastically, as new facts can easily extend the

knowledge base as do new rules or relations. Lastly,

every decision that such a system makes can be easily

explained to a human user by extracting the chain of

rules and facts which led to an inference.

2.1 Semantic Maps

Enriching spatial maps for robots goes back at least

to the works of Kuipers, Buschka in the early 00s

(Kuipers, 2000; Buschka, 2005). Galindo et al sug-

gest a hierarchy of map representations (spatial, se-

mantic) and ”anchor” semantic meaning to objects

and places (Galindo et al., 2005). N

uchter and

Hertzberg showed how semantic annotation can be

automatically applied to 3D SLAM Maps by reason-

ing over previous knowledge in form of a constraint

network to enhance scene interpretation and automate

annotation (N

uchter and Hertzberg, 2008). Pronobis

reﬂects on place classiﬁcation and the generation of

semantic maps and the enhancement of object search.

The mapping process is deﬁned broadly as associat-

ing spatial concepts (e.g. ”kitchen”) with spatial enti-

ties (e.g. a polygon on map) (Pronobis, 2011). Lang

and Paulus formally deﬁne a semantic map as a hy-

brid map of spatial and semantic information with the

stress of the semantic information being represented

in a way which allows for inference (Lang and Paulus,

2014). At the same time Kunze et al use reasoning to

enhance object recognition (Kunze et al., 2014). In

a survey Kostavelis and Gasteratos claim that most

works regarding semantic mapping describe robotic

indoor scenarios, note however that the use of seman-

tic maps for robotic applications increases (Kostavelis

and Gasteratos, 2015). Deeken et al generate low

level occupancy grids from a semantic map (Deeken

et al., 2015) while other works apply semantic maps

to navigation in form of land mark recognition (Cos-

gun and Christensen, 2018). In recent years, Kunze et

al anticipate a strong yet growing trend for semantic

Online Inference of Robot Navigation Parameters from a Semantic Map

157

Rule-Based Inference System /

Reasoner

Domain-Specific Rules

(Extended)

SPARQL-Interface

Entity

Component Component ...

Persistence

Figure 1: Overview of the system architecture of SEMPR.

maps for long term autonomy in their survey (Kunze

et al., 2018). A rule based reasoner is used by Deeken

et al to infer process phases in agricultural machines

(Deeken et al., 2019). Recently, Crespo et al surveyed

a growing number of approaches to semantic naviga-

tion by using semantic maps with reasoning methods.

The focus, however, still lies on using semantic in-

formation for human-robot-interaction and exploiting

high level information for room recognition and clas-

siﬁcation (Crespo et al., 2020).

In this work we use rule based reasoning over a se-

mantic map of an agricultural deployment site to infer

not where the robot should go but how to act on the

way. Like Deeken et al generated occupancy grids

(Deeken et al., 2015), in our work we generate traver-

sal costs for the whole deployment site which enables

a more precise approach to path planning. In contrast

to the descriptions of indoor scenarios the focus rests

not on objects, but on the properties of free space ar-

eas.

3 SYSTEM DESIGN

3.1 Inferring Robot Control Parameters

The framework we developed to hold the representa-

tion and infer implications for robot control and navi-

gation is called Semantic Environment Mapping, Pro-

cessing and Reasoning (SEMPR)

and added as a li-

brary to the robots system.

Its Architecture, also shown in Fig. 1, comprises

of a collection of Entities which make up the known

facts about the environment, the domain speciﬁc rules

which deﬁne how to infer new knowledge and a

reasoning system to perform the automated infer-

ence. Interfaces to a persistence layer and a SPARQL-

Query-Service enable further usability.

The Entities themselves are made up by a set of

components which encode arbitrary data like geome-

tries, semantic information or transformations. How-

https://github.com/sempr-tk/sempr

ever, all semantic information is described according

to the Resource Description Framework which is a

well known format for machine understanding and

reasoning (W3C, 2004).

The Rules facilitate inference in the form if-then.

The inferred information is added to the knowledge

base and can be used to activate robot behaviour,

set control parameters or be used in chained rules.

Knowledge of different types can be combined to en-

able reasoning over geospatial and semantic informa-

tion alike as seen in the example below:

[robotInZone:

(?robot <type> <Robot>),

(?zone <type> <Zone>),

Geometry(?zone ?zoneGeo),

Geometry(?robot ?robotGeo),

geo:intersects(?zoneGeo ?robotGeo)

(?robot <inZone> ?zone)]

Which translates to For each entity R ∈ Robots

and each entity Z ∈ Zones, retrieve the respective ge-

ometries G

, G

. For each combination where G

in-

tersects G

, add the fact that R is in zone Z.

3.1.1 Reasoner

In order to decide which rules need to be triggered

at a time, the inference system makes use of the

RETE pattern matching algorithm (Forgy, 1982): The

textual representation of the rules is parsed into a

graph where every node implements a small check

on the given data. If the check succeeds, the data

is forwarded to its child nodes. The connections be-

tween the nodes thus construct the complex condi-

tions as stated in the rules, while the terminal nodes

implement the rules’ consequences. By also insert-

ing memory-nodes, the pattern matching implements

a trade-off between performance and memory con-

sumption, as they enable an iterative processing of

changes to the knowledge base. Facts that match a set

of conditions are stored in the memory nodes and can

be used as partial matches in subsequent conditions

and effects without re-evaluation. Furthermore, when

retracting facts, only branches with memory nodes

containing them need to be re-evaluated. The whole

graph consists of two parts: The alpha-network, in

which the basic data elements are inspected indepen-

dently of each other, and the beta-network, in which

multiple conditions get combined and more complex

checks can be performed.

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

158

3.2 Robot Control Architecture

For our setup we implement a ROS node called

sempr ros bridge from which a SEMPR instance

is created and managed. The semantic map

and rules are loaded from pregenerated ﬁles on

startup. Fig. 2 shows, how the navigation modules

move base flex and waypoint server are con-

nected. The sempr ros bridge node polls updates

from the SEMPR instance in a speciﬁed time interval

via the SPARQL Protocol And RDF Query Language

(SPARQL) interface. The resulting information is then

converted into typical ROS message formats and ei-

ther published or introduced to the system in form of

online reconﬁguration requests. Additional informa-

tion that the robot could not sense on its own, such

as weather data, can be entered manually directly into

the knowledge base to test the system.

Figure 2: Integration of the reasoning module into a ROS

based system.

3.3 Semantic Parameters for Robot

Control

In our work we concentrated on some central aspects

of robot control and navigation to be inferred by the

rule engine.

3.3.1 Global Path Costs

One very well researched aspect of mobile robots is

the problem of planning a path from one spatial lo-

cation to another in an optimal manner according to

certain variables. Common criteria to be optimized by

a path planning algorithm are the length of the path,

power consumption and the time of traversal, but de-

pending on the domain other considerations might be

of interest: In agriculture, the distribution of strain

to the soil by heavy machinery can be very impor-

tant, as well as to route trafﬁc in a speciﬁc way to

ascertain smooth workﬂows. In robotic practice, dif-

ferent surface properties can have a huge impact on

the drivability of the area. A common way to model

these is the utilization of costmaps: Different cost val-

ues are assigned to discrete grid cells of an area as

a function of the optimization criteria. By assigning

different cost values to discrete grid cells of an area

as a function of the optimization criteria, a shortest

path planning algorithm such as A* or Dijkstra’s can

easily optimize for lowest overall path costs instead

of geometric length only. In that way, more costly

areas such as grass or soil are only crossed by the

robot if the routes are signiﬁcantly shorter than routes

only traversing better suited surfaces like pavement.

Naively, it is possible to assign traversal costs in this

way to all areas of the map and thus model the de-

ployment environment of the robot. However, such a

static approach meets its limits as soon as the input

parameters to the cost function change. For exam-

ple, the drivability of unsealed surfaces such as soil

or grassland change drastically depending on whether

the surface is wet or dry. By modelling static knowl-

edge about the surface properties of the environment

and according rules, we can infer the costs to tra-

verse a certain area during runtime and thus gener-

ate a costmap dynamically with the logical rules mod-

elling the function to calculate the costs of traversing

a given area given the input variables at the time. In

our experiment we deﬁned traversal costs for discrete

values of very low, low, medium, high and very high.

Those values then would be assigned to regions of the

map according to rules, taking into account the sur-

face type and overall ground wetness to demonstrate

the dynamic generation of a 2D costmap.

3.3.2 Fences and Gates

On a farm where animals are kept and moved between

different places, fences and gates play a large role.

While only few gates can be opened by the robot it-

self right now, it is very plausible that there might

be automated systems opening the gates mechanically

integrated in the future. At the moment, however, it

would already be useful if the robot knew which gates

were open or closed at a given time. In practice some

gates’ open/closed state or probabilities thereof can in

fact be derived from certain knowledge like the time

of day or information about the whereabouts of the

animals. If a gate is known or reasoned to be open

it can be represented in the static map layout, which

updates whenever the state of a gate changes.

3.3.3 Local Path Planning Algorithms and

Driving Speeds

Typical navigation stacks found in robots based on

ROS use a local planning module in addition to the

global path planning. It takes into account locally

observed obstacles, the kinematics of the robot as

well as strategies to account for smooth motor con-

trol. State of the art are multiple approaches of high

quality local planners such as dwa (Fox et al., 1997),

Online Inference of Robot Navigation Parameters from a Semantic Map

159

teb (Roesmann et al., 2012) and eband (Quinlan and

Khatib, 1993). However, experience shows that dif-

ferent planners perform differently in different envi-

ronments. So, in a heterogeneous environment it is

obviously best to choose the planner depending on the

local surroundings. To enable this choice, we mod-

elled the criteria of dynamics which we deﬁne as a

scalar between 0.0 and 1.0 to represent the presence of

moving objects in the area the robot is going to cross.

On the extreme value of 1.0, possibly contrary to intu-

ition, it can make sense to choose an algorithm which

sticks closer to the globally planned path and resolves

situations of being blocked by waiting or reevaluation

of the obstacle map instead of trying to locally evade

obstacles which possibly move quicker than the robot

itself. A dynamic value of 0.0 means the robot can

expect to be the only moving object in the vicinity,

which in turn means every obstacle sensed will prob-

ably be static and should be smoothly evaded using a

ﬁtting local planning algorithm.

As a second input parameter we deﬁned the

freespace value to represent the structuredness and

openness of an area. For robotic practice a wide open

ﬁeld with few to no obstacles blocking sight means

that a local planner can optimize for smooth move-

ments freely without taking moving obstacles into ac-

count that could appear behind occlusions suddenly.

Vice versa, a cluttered environment means that the

robot should try to stick closely to the global path and

possibly reduce driving speeds to accommodate the

fact that obstacles might only become visible when

already close to the robot. These control parameters,

like maximum velocity and the used planning algo-

rithm, can be inferred from the semantic state of the

environment for the different zones with respect to

their aforementioned properties. Their actual geomet-

ric shape and the robots location can also be taken

into account during the reasoning process, as shown

earlier.

3.4 Modelling of the Semantic Map

To model such a semantic map we model the knowl-

edge as a collection of facts about the domain and a

set of rules which deﬁnes allowed inferences. On a

symbolic level this means to have a set of entities rep-

resenting physical or virtual objects of the environ-

ment. Those are associated primarily with geospatial

geometric information as a lot of relevant information

in the agricultural domain revolves around spatial lo-

cations, areas and their extents. We used the open

source tool QGIS

to model zones as geospatial vector

polygons in ESRI shape-layers shown in Fig. 3. Each

www.qgis.org

Figure 3: Zones as modelled in QGIS.

Table 1: Example data after the postprocessing steps.

groundType

freespace

dynamic

animals

1 gravel 0.25 0.4 false

2 gravel 0.7 0.8 false

3 gravel 0.8 0.75 true

4 gravel 0.35 0.2 false

5 gravel 0.15 0.2 false

6 gravel 0.6 0.75 true

zone constitutes a primary semantic entity. The ﬁle

format allows to annotate a layer with a table as seen

in Tab. 1 with each row representing a zone. These

are used to model semantic facts which make up RDF

triples with the row as subject, the column name as

predicate and the value of the cell corresponding to

row + column as object. Rules are deﬁned as a pair

of two lists: One for the preconditions and a second

for the assertions. The rules are then stored in a sep-

arate, human readable text ﬁle. The shapes and rules

are then imported into the semantic representation and

reasoning module described in section 3.1.

As mentioned before in 3.1, the rules can be con-

sidered to represent logical functions of the input val-

ues which means they can be derived from expert

knowledge using conventional acquisition techniques

like case studies, simulations or structured and un-

structured interviews. An advantage of using a rule

based approach can be seen in the fact that rules can

be naturally formulated as in “traversal costs increase

for certain surface types when the individual wetness

threshold is reached” and then easily translated into

the machine-readable form.

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

160

4 RESULTS

Fig. 4 shows two costmaps generated online by the

system. On the top is a generated costmap of the ar-

eas if the condition dry is met as a global condition.

On the bottom a costmap is generated for a moist con-

dition met globally.

Figure 4: Inferred Costmaps: blue, purple, red, turquoise

(low – high). Top: Costs for dry farm condition. Bottom:

Costs for wet farm condition.

As can be seen in ﬁg. 5 the generated costs can be

incorporated in the global costmap of the ros naviga-

tion system and used online for path planning. This

way the navigation behaviour of the robot can be

adapted online according to updates from the reason-

ing process.

Furthermore, the navigation stack of this robotic

system is able to switch between different local plan-

ning algorithms online which additionally offer an

interface to update certain parameters such as maxi-

mum speed and acceleration online via dynamic re-

conﬁgure. For example, the data used by the con-

text navigation algorithm shown in (Hoellmann et al.,

2020) can be generated by the semantic control setup.

This way it is possible for the robot to directly infer

important information dynamically during its deploy-

ment. The ability to generate costmaps dynamically

and thus react to changes in weather or other condi-

Figure 5: Robot planning a path according to costmap.

tions can represent real world conditions much more

accurately than any static approach. Likewise can the

inference of choices like local planning algorithms or

movement parameters open up new approaches of in-

telligent robot behaviour. To use a semantic reason-

ing system instead of implicit knowledge encoded in

the programming of the robot allows to identify rules

and facts with the ﬁelds respective experts using nat-

urally formulated rule syntax. Decisions made by the

robotic system in this way can be easily explained

and documented due to the explicit modelling of the

knowledge base.

5 OUTLOOK

We showed how a semantic map, that is a represen-

tation of a geographical area with georeferenced sub-

areas, can be annotated in a way to allow efﬁcient in-

ference using known tools like QGIS together with

a reasoning system for geospatial and semantic in-

formation based on the well known RETE algorithm.

With the integration into the ROS ecosystem, this can

be used for the online change of parameters essen-

tial for robotic applications, thus increasing the ﬂex-

ibility with which a robot can adapt to changes in a

complex environment like agriculture. Relevant rules

and facts can be deﬁned using conventional knowl-

edge acquisition techniques with experts. In agricul-

ture it makes sense to organize those rules and facts

around geospatial information, thus being easily ap-

proachable to discuss with experts like farmers.

A further advantage of such a system is to deﬁne

knowledge in an explicit way, enabling the robot to

document an explanation for each of its decisions and

proﬁting of sensing information as well as of prior

knowledge and even inferring from knowledge that

cannot or will not be sensed by the robot.

Future work will include the integration of the in-

ferred information in a live robotic system, replacing

Online Inference of Robot Navigation Parameters from a Semantic Map

161

existing solutions to look up context based control pa-

rameters with the demonstrated inference system. It

appears promising to extend the control parameters

to infer, e.g., covariances for localization ﬁlters de-

pending on the traction to expect on certain surfaces

or additional behavioral strategies like acoustic or op-

tical signalling when expecting to act in the vicinity

of humans or animals. Also strategies like not opti-

mizing for shortest paths but instead following right

hand rules when driving along paths or streets might

be beneﬁcial. In a wider perspective the extension

towards probabilistic reasoning appears sensible to

cope with information not easily conveyed with sim-

ple facts.

The generation of ﬁne-grained navigation

costmaps provides the foundation for further work:

As a future avenue we plan to use the semantic

representation to map detected obstacles and annotate

additional information. In the long term there are

many kinds of obstacles which might move in the

scale of minutes, hours or sometimes days. Instead

of just adding them to the costmap and remove them,

once not seen anymore, it might be better to actively

check once a certain amount of time has passed or

not regard them for path planning.

Using semantic reasoning technologies can be an

important contribution to add to the ﬂexibility and

thus robustness of robots expected to act in complex

environments without human supervision. Making

knowledge about the environment explicit can add

to the explainability of artiﬁcial intelligent decisions

made by robots as well as to the ease in identifying

relevant rules and facts with the help of experts in the

respective ﬁeld like agriculture.

ACKNOWLEDGEMENTS

The DFKI Niedersachsen Lab (DFKI NI) is spon-

sored by the Ministry of Science and Culture of

Lower Saxony and the VolkswagenStiftung. The

paper describes work that has been developed in

the context of the funded projects Experimentier-

feld Agro-Nordwest (BMEL, 28DE103E18), DAKIS

(BMBF, 031B0729B) and ZLA (NiMWK, Volkswa-

genstiftung, ZDIN 11-76251-14-3/19).

REFERENCES

Bergerman, M., Billingsley, J., Reid, J., and Henten, E. V.

(2016). Robotics in Agriculture and Forestry. In Sicil-

iano, B. and Khatib, O., editors, Springer Handbook

of Robotics, chapter 56, pages 1463–1492. Springer.

Buschka, P. (2005). An investigation of hybrid maps for

mobile robots. PhD thesis, University

Orebro.

Cosgun, A. and Christensen, H. I. (2018). Context-aware

robot navigation using interactively built semantic

maps. Paladyn, Journal of Behavioral Robotics,

9(1):254–276.

Crespo, J., Castillo, J. C., Mozos, O. M., and Barber, R.

(2020). Semantic Information for Robot Navigation:

A Survey. Applied Sciences, 10(2):497.

Deeken, H., Wiemann, T., and Hertzberg, J. (2019). A

spatio-semantic approach to reasoning about agricul-

tural processes. Applied Intelligence, 49(11):3821–

3833.

Deeken, H., Wiemann, T., Lingemann, K., and Hertzberg,

J. (2015). SEMAP - a semantic environment mapping

framework. In 2015 European Conference on Mobile

Robots (ECMR), pages 1–6. IEEE.

Egerstedt, M., Pauli, J. N., Notomista, G., and Hutchinson,

S. (2018). Robot ecology: Constraint-based control

design for long duration autonomy. Annual Reviews

in Control, 46:1–7.

Forgy, C. L. (1982). Rete: A fast algorithm for the many

pattern/many object pattern match problem. Artiﬁcial

Intelligence, 19(1):17–37.

Fox, D., Burgard, W., and Thrun, S. (1997). The dy-

namic window approach to collision avoidance. IEEE

Robotics Automation Magazine, 4(1):23–33.

Galindo, C., Safﬁotti, A., Coradeschi, S., Buschka, P.,

Fernandez-Madrigal, J., and Gonzalez, J. (2005).

Multi-hierarchical semantic maps for mobile robotics.

In 2005 IEEE/RSJ International Conference on Intel-

ligent Robots and Systems, pages 2278–2283. IEEE.

Hoellmann, M., Kisliuk, B., Krause, J. C., Tieben, C.,

Mock, A., Sebastian, P., Igelbrink, F., Wiemann,

T., Martinez, S. F., Stiene, S., and Hertzberg, J.

(2020). Towards Context-Aware Navigation for Long-

Term Autonomy in Agricultural Environments. In

2020 IEEE/RSJ International Conference on Intelli-

gent Robots and Systems (IROS), Las Vegas.

Kostavelis, I. and Gasteratos, A. (2015). Semantic mapping

for mobile robotics tasks: A survey. Robotics and

Autonomous Systems, 66:86–103.

Kuipers, B. (2000). The Spatial Semantic Hierarchy. Artiﬁ-

cial Intelligence, 119(1-2):191–233.

Kunze, L., Burbridge, C., Alberti, M., Thippur, A.,

Folkesson, J., Jensfelt, P., and Hawes, N. (2014).

Combining top-down spatial reasoning and bottom-

up object class recognition for scene understanding.

In 2014 IEEE/RSJ International Conference on Intel-

ligent Robots and Systems, pages 2910–2915. IEEE.

Kunze, L., Hawes, N., Duckett, T., Hanheide, M., and Kra-

jnik, T. (2018). Artiﬁcial Intelligence for Long-Term

Robot Autonomy: A Survey. IEEE Robotics and Au-

tomation Letters, 3(4):4023–4030.

Lang, D. and Paulus, D. (2014). Semantic Maps for

Robotics. In Proceedings of the Workshop on AI

Robotics at the IEEE International Conference on

Robotics and Automation.

Langley, P., Meadows, B., and Sridharan, M. (2017). Ex-

plainable agency for intelligent autonomous systems.

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

162

In IAAI ’17 proceedings of the twenty-ninth AAAI con-

ference on innovative applications of artiﬁcial intelli-

gence, pages 4762–4763.

uchter, A. and Hertzberg, J. (2008). Towards semantic

maps for mobile robots. Robotics and Autonomous

Systems, 56(11):915–926.

Pronobis, A. (2011). Semantic Mapping with mobile robots.

PhD thesis, KTH.

Putz, S., Santos Simon, J., and Hertzberg, J. (2018). Move

Base Flex A Highly Flexible Navigation Framework

for Mobile Robots. In 2018 IEEE/RSJ International

Conference on Intelligent Robots and Systems (IROS),

pages 3416–3421. IEEE.

Quinlan, S. and Khatib, O. (1993). Elastic bands: con-

necting path planning and control. In Proceedings

IEEE International Conference on Robotics and Au-

tomation, pages 802–807 vol.2.

Roesmann, C., Feiten, W., Woesch, T., Hoffmann, F., and

Bertram, T. (2012). Trajectory modiﬁcation con-

sidering dynamic constraints of autonomous robots.

In ROBOTIK 2012; 7th German Conference on

Robotics, page 74–79.

Shamshiri, R. R., Weltzien, C., A. Hameed, I., J. Yule, I.,

E. Grift, T., K. Balasundram, S., Pitonakova, L., Ah-

mad, D., and Chowdhary, G. (2018). Research and

development in agricultural robotics: A perspective of

digital farming. International Journal of Agricultural

and Biological Engineering, 11(4):1–11.

W3C (2004). Resource Description Framework. https://

www.w3.org/RDF/.

Online Inference of Robot Navigation Parameters from a Semantic Map

163