Coresense D1.3

CS-050-v1.
0 HORIZON
2023/11/05 EUROPE
Public
D1.3
Theory of
Understanding
An initial formal definition
Ricardo Sanz (UPM), Manuel Rodriguez (UPM), Martin

Molina (UPM), and Esther Aguado (UPM)
CORESENSE
Horizon Europe Project #101070254
HORIZON-CL4-2021-DIGITAL-EMERGING-01-11
The CORESENSE Project is a four year research project focused in the development of a technology of machine
understanding for the construction of dependable autonomous systems. It is a project funded by the European
Union.
Programme: Horizon Europe
Project number: 101070254
Project name: CoreSense: A Hybrid Cognitive Architecture for Deep Understanding
Project acronym: CORESENSE
Topic: HORIZON-CL4-2021-DIGITAL-EMERGING-01-11
Type of action: HORIZON Research and Innovation Actions
Granting authority: European Commission-EU
Project start date: 1 October 2022
Project end date: 30 September 2026
Project duration: 48 months
Most of CORESENSE results are open science but IPR regulations may still hold. Requests for permission to
reproduce this document or to prepare derivative works of this document should be addressed to the CORE-
SENSE Consortium IPR management.
Copyright © 2022-2026 by The CORESENSE Consortium.
UNIVERSIDAD POLITECNICA DE MADRID (UPM)
TECHNISCHE UNIVERSITEIT DELFT (TUD)
FRAUNHOFER GESELLSCHAFT ZUR FORDERUNG DER ANGEWANDTEN FORSCHUNG (FHG)
UNIVERSIDAD REY JUAN CARLOS (URJC)
PAL ROBOTICS SL (PAL)
IRISH MANUFACTURING RESEARCH COMPANY LIMITED BY GUARANTEE (IMR)
CZECH TECHNICAL UNIVERSITY IN PRAGE (CVUT)
For information about other CORESENSE products, reports or articles, please visit our Web site:
http://www.coresense.eu/
NO WARRANTY — Please read the fine print:

THIS CORESENSE MATERIAL IS FURNISHED ON AN “AS-IS” BASIS. THE CORESENSE CONSORTIUM NOR ANY OF THE INDIVIDUAL CORE-
SENSE PARTNERS OR RESEARCHERS MAKE NO WARRANTIES OF ANY KIND, EITHER EXPRESSED OR IMPLIED, AS TO ANY MATTER INCLUD-
ING, BUT NOT LIMITED TO, ACCURACY, COMPLETENESS, WARRANTY OF FITNESS FOR PURPOSE OR MERCHANTABILITY, EXCLUSIVITY,
OR RESULTS OBTAINED FROM USE OF THE MATERIAL. CORESENSE DOES NOT MAKE ANY WARRANTY OF ANY KIND WITH RESPECT TO
FREEDOM FROM PATENT, TRADEMARK, OR COPYRIGHT INFRINGEMENT.
The CORESENSE project is funded by the EC Horizon Europe programme though grant HE
#101070254 inside the HORIZON-CL4-2021-DIGITAL-EMERGING-01-11 topic. Views and
opinions expressed in this document are however those of the author(s) only and do not
necessarily reflect those of the European Union or the Horizon Europe Programme. Nei-
ther the European Union nor the granting authority can be held responsible for them.
Executive Summary
Title: D1.3 Theory of Understanding

Subtitle: An initial formal definition
Author: Ricardo Sanz (UPM), Manuel Rodriguez (UPM), Martin Molina (UPM), and Esther Aguado
(UPM)
Reference: CS-050 v 1.0
Date: 2023/11/05
Dissemination level: Public
URL: http://www.coresense.eu/doc/CS-050.pdf
This is CORESENSE Deliverable 1.3. It contains an initial attempt at a formalisation of the

CORESENSE theory of understanding, the central object of development in the project. This
deliverable describes only the initial work of the project on the theory. The project theoret-
ical developments will continue throughout the project until the release of the final theory
deliverables (D1.3v2, D1.4v2, and D9.7) at the very end of the project.
The structure of this document is as follows:
Chapter 1 contains an introduction to the theme and the document.
Chapter 2 discusses the strategy for the elaboration of the theory.
Chapter 3 proposes the initial snapshot of the theory.
Chapter 4 performs an initial analysis of the theory.
Annex 1 presents an initial basic reference of category theory.
CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05 | 3-49

Document Versions
Version 0.1 | 2023/09/01

Changes Initial draft
Authors R.Sanz
Version 0.2 | 2023/10/20

Changes Release candidate
Authors R.Sanz
Version 0.3 | 2023/10/20

Changes Corrections after QS
Authors R.Sanz, M. Molina, N, Hammoudeh, E. Aguado, M. Rodríguez
Version 1.0 | 2023/11/05

Changes Final release
Authors R.Sanz
Deliverable Approval
2023/10/22 — WPL: M.Rodriguez (UPM)
2023/10/20 — PC: R.Sanz (UPM)
2023/11/03 — QA: N.Hammoudeh (FHG)
4-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

Contents
Executive summary 3
Document Versions 4
Contents 5
List of Figures 7
List of Tables 9
1 Introduction 11
1.1 Towards Understanding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.2 Context and Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.3 Document Structure and Content . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2 Strategy for building the theory 15

2.1 Understanding in CORESENSE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 On Theories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2.1 Characteristics of a Scientific Theory . . . . . . . . . . . . . . . . . . . . . . 16
2.2.2 Operationalisation of a Scientific Theory . . . . . . . . . . . . . . . . . . . 18
2.2.3 Essential Elements for a Theory of Understanding . . . . . . . . . . . . . . 18
2.3 Steps for a Theory of Understanding in CORESENSE . . . . . . . . . . . . . . . . . 20
3 Towards a Theory of Understanding 23

3.1 Building blocks for a Definition of Understanding . . . . . . . . . . . . . . . . . . 23
3.2 A definition of understanding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.2.1 Degrees of understanding . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.2.2 Some considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4 Analysis 31
4.1 Understanding and Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5
Contents
4.2 Coverage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.2.1 Coverage of perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.2.2 Coverage of the Testbeds’ Needs . . . . . . . . . . . . . . . . . . . . . . . 33
4.3 Future development of the theory . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.3.1 Other directions of development . . . . . . . . . . . . . . . . . . . . . . . 39
4.3.2 Reification as reusable assets . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.3.3 Towards a CT Definition of Understanding . . . . . . . . . . . . . . . . . . 39
Bibliography 43
A Category Theory for Roboticists 47

A.1 Category . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
A.2 Functor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
A.3 Natural Transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
6-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

List of Figures
1.1 Cognitive science sits at the convergence point of many disciplines, of both hu-
man and technical nature. Our theory of understanding targets the general do-
main with cross-cutting implications in subdomains. . . . . . . . . . . . . . . . . . 12
2.1 Theory development proceeds along the project in a feedback relation with testbed
development. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.1 A general open system, it is composed of different subsystems (system of systems). 24

3.2 A phenomenon (in red) as an instance of a system. . . . . . . . . . . . . . . . . . . 25
3.3 Left figure shows partial understanding of the phenomena as there is no model
of the interaction between them, the right figure shows a model covering also
the interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.1 While ”understanding” is commonly tied to language processing the deep un-
derstanding that we need for our robots is not just language, it is systemic: the
cognitive agent creates an abstract system in its head that matches some spe-
cific aspect of the reality. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
7
List of Figures
8-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

List of Tables
2.1 Essential Elements for a Theory of Understanding . . . . . . . . . . . . . . . . . . 19
4.1 Coverage of perspectives of understanding . . . . . . . . . . . . . . . . . . . . . . 33
9
List of Tables
10-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

1 Introduction
”If the organism carries a ‘small-scale model’ of external reality

and of its own possible actions within its head, it is able to try out
various alternatives, conclude which is the best of them, react to
future situations before they arise, utilise the knowledge of past
events in dealing with the present and future, and in every way
to react in a much fuller, safer, and more competent manner to
the emergencies which face it.” [Craik, 1943, p. 61]
The question of ”understanding” is a central problem in cognitive science. However, with the
relevant exceptions of language and education, ”understanding” has mainly stayed as a mere
descriptive term or a topic for abstract philosophical debate. Only recently has it received at-
tention from the artificial intelligence community as a core issue for intelligent machines. A
similar analysis can be done concerning the question of ”awareness”, with the extra difficul-
ties associated to the analysis of phenomenological experience and its potential realisation in
machines.
The problems of understanding and awareness by machines are of special importance in the
case of autonomous robots because misunderstandings concerning what is going on in the
robot environment may lead to catastrophic failures in real-world deployments. Being aware of
the world and itself, understanding of orders, understanding its environment, understanding
itself, understanding others —esp. humans—, become critical capabilities for the autonomous
robot to properly complete a mission in an open, dynamic and uncertain environment.
So far, these problems have been tackled by directly engineering specific ad-hoc solutions for
concrete deployed systems. However, there is yet no solid, universally accepted ”theory of un-
derstanding and awareness” that can guide, in general, the construction of cognitive architec-
tures for autonomous robots. Only in the domain of artificial general intelligence is it possible
to find some initial attempts at such a theory.
1.1 Towards Understanding

The CORESENSE Project1 seeks to improve the capability of understanding of autonomous
robots. It tries to do so using a rigorous engineering process based on a explicit Theory of
Understanding.
Given the nature of the project, robotics and artificial intelligence will be the particular foci
domains of interest, but we also intend a more general cognition approach to defining ”un-
derstanding”, with a clear ambition of universality. Our desired definition and its associated
theory shall be applicable to any cognisant system that understands.
1
http://coresense.eu
11
Chapter 1. Introduction
Psychology
Linguistics Philosopy
Cognitive
Science
Artificial
Understanding Education
Intelligence
Automatic
Robotics
Control
Cybernetics
CS-IMG-023-U
Figure 1.1: Cognitive science sits at the convergence point

of many disciplines, of both human and technical nature.
Our theory of understanding targets the general domain
with cross-cutting implications in subdomains.
We will not include here all the relevant text from the state-of-the-art analysis [Sanz et al., 2023],
but some of its content will be repeated here to obtain a more or less self-contained document.
The concept of understanding is a fishy target. A common view of understanding is being able
to label things in scenes: ”Car, car, car, person, traffic light”, and we can hear the environment
understanding mechanisms of an autonomous car. Some tests implicitly assert that under-
standing is the capability of responding adequately to given orders: ”Robot, bring me a glass
of water”, and the robot performs as ordered. Some authors argue that understanding means
being able to explain. ”Robot, why did you put this glass of water on the table?” and the robot
tells us that it did so due to following our orders. Other people do think that understanding is
getting useful mental representations, or reaching agreements with people and the world, or
answering questions about stories, or being aware of owned knowledge. There are so many
understandings of understanding [Mason, 2003, Baumberger, 2014].
However, what guides us in the CORESENSE project, and provides a cornerstone for the con-
tent of this document —a theory of understanding— is that we believe that:
• There is some core commonality behind all those situations because all of them show
a capability of using some information meaningfully, hence implying that there exists a
fundamental aspect of understanding that powers all of them; item that this core com-
ponent is essential for achieving higher levels of capability, performance, robustness and
resilience; and
• that such core interpretation of understanding can be captured using systematic and rig-
12-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

1.2. Context and Motivation
orous methods from systems theory.
This document will try to provide such a definition after distilling our understanding of under-
standing as a collective that was the content of [Sanz et al., 2023]. We will propose an oper-
ational synthesis of all of them to address the needs of both robot engineers and cognitive
scientists at large.
1.2 Context and Motivation

In CORESENSE, we seek a theory and a derived technology to make machines that understand
more, or better, or maybe just enough to perform their task sufficiently well. Understanding
seems central to the human ability to be aware of what change is necessary to bring forth in
particular contexts or during the performance of particular tasks. For autonomous robots, it
shall be the same.
It could well be the case that ”machine understanding” is just a red herring that takes us out of
the right path to trustworthy machines. ”Machines lack semantic capability”, said John Searle
in the 80’s, and the echoes of this assertion are still around us. Maybe this is the case, and
maybe it will ever be.
However, it is pretty obvious that when the thermostat uses a bimetal to sense the room tem-
perature and switches off a relay of a heater the machine understands that the room is too
hot compared to what I told it. At least to the extent of being able to do what I asked it for.
Also, it is pretty obvious that when I say ”Hey Google, Call Mom” to the Google assistant in my
cellphone and it places a phone call to my mother the machine understood me. At least to
the extent of being able to do what I asked it for.
Appropriate overt behaviour in response to utterances is certainly a proof of language under-
standing, but it is not always a sine qua non condition [Ziff, 1972]. Understanding orders com-
ing from human masters is certainly an important aspect of understanding, but in autonomous
machines there are other forms of understanding that are also of major value: understanding
road signs; understanding human emotion; understanding weather effect; understanding hu-
man needs beyond the immediate orders; understanding that with a broken leg a bipedal robot
cannot walk as planned, etc.. The pursuit of a theory of understanding is daunting: There are so
many relevant objects of understanding for so many understanding subjects in so many contexts
of understanding. Can a single theory/technology address all of them?
The idea of understanding sits at the heart of cognitive science (see Figure 1.1). However, we
have not reached yet an agreement on what it is with precision enough as to operationalise it
in the technical domains of intelligent artefacts. Clarifying the concept will be a major aspect
of the pending work.
It is necessary to be precise about the specific stance that marks this endeavour. In this project
we try to elucidate the nature of understanding from the concrete stances of 1) autonomous
cognitive systems, 2) autonomous cognitive systems builders and 3) autonomous cognitive sys-
tems users.
In this report we unify these three into a specific systems engineering perspective2 . From this
perspective, we analyse extant views on understanding and propose a concrete understanding
of understanding to be reified in the core architectural mechanisms of CoreSense.
This report does not propose a new cognitive architecture targeting understanding, nor better
representations, programming languages, algorithms, or methods relevant to understanding
2
A systems engineering perspective aggregates multiple stakeholders views into a single, coherent picture.
CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05 | 13-49

Chapter 1. Introduction
and its mechanisation. The report focuses on a conceptual analysis and the introduction of a
new theoretical approach to understanding —the capability of envisioning of future value— and
some architectural guidelines for AIs to implement relevant mechanisms for improved under-
standing of situations that serve as a basement for robot autonomy and even future machine
consciousness.
1.3 Document Structure and Content

This is the CORESENSE Deliverable D1.3 and contains a Formalization of the theory of under-
standing as stated in the GA [CORESENSE Consortium and EC, 2022].
This deliverable presents the initial efforts concerning the theory and will be re-released at the
end of the project in M48.
The structure of this document is simple:
Chapter 1: This chapter sets the context for the rest of the document.
Chapter 2: Describes the strategy to build the theory.
Chapter 3: Is the main content of the document and provides the current snapshot of the
theory based on set theory.
Chapter 4: Analyses the current version (1.0) of the theory from the stated needs and consid-
ers its potential future development using category theory.
Annex: Provides a brief introduction to category theory.
14-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

2 Strategy for building the theory
A theory is a rational type of abstract thinking about a phe-

nomenon or the results of such thinking. The process of con-
templative and rational thinking is often associated with such
processes as observational study or research. Theories may be
scientific, belong to a nonscientific discipline, or no discipline at
all. …In modern science, the term ”theory” refers to scientific
theories, a well-confirmed type of explanation of nature, made
in a way consistent with the scientific method and fulfilling the
criteria required by modern science. Such theories are described
in such a way that scientific tests should be able to provide em-
pirical support for it, or empirical contradiction (”falsify”) of it.
[Wikipedia contributors, 2023]
The main content of this deliverable (D1.3) is an initial presentation of the theory of understand-
ing as stated in the GA [CORESENSE Consortium and EC, 2022]. Building such a theory is going
to proceed in stages. This deliverable presents just a first release towards the final formal ex-
pression of such a theory. As stated in the GA [CORESENSE Consortium and EC, 2022], it will
be re-released in final form at the end of the project (in M48). We also expect to have non re-
leased versions of the theory during project execution. Interim releases will be made available
through the project website and the Open Science repositories.
The content of this chapter is a description of the overall strategy of the project towards the
development of the theory.
2.1 Understanding in CORESENSE

The CORESENSE proposal offered an initial glimpse at the idea of understanding that we are
pursuing in the project.
”the capability of understanding will be based on the agent having actionable, hybrid,
causal models, that can be populated, adapted, exploited, and shared with other agents”
In deliverable D1.1 [Sanz et al., 2023] we analyse some of the concepts of understanding that
have been proposed in the past in the broad domain of cognitive systems. In most of the cases
they were part of linguistic theories. In a small number of cases, they are related to Artificial
Intelligence (AI) and robotics One possible strategy was to attempt a synthesis of the most
relevant theories from the perspective of the needs of the CoreSense architecture.
For example, Zagzebski [Zagzebski, 2019] proposed that understanding is the grasp of struc-
15
Chapter 2. Strategy for building the theory
ture. The structure of an object shapes it, gives it organization and unity and this lets us see it
as a single object. When we grasp an object’s structure, we understand the object. This idea
of understanding suits ours, because our models are simplified representations of what the
agent grasps. The larger and more complex the object of understanding, the bigger the sim-
plification that may leave out of the model object aspects that may be important at different
times or for different purposes.
The synthesis is a good strategy to get at results of broader acceptance. However, in the project
we shall also follow a more pragmatic approach to understanding: it shall be useful in the de-
velopment of intelligent machines. At the end, we need an hybrid approach. The CORESENSE
theory of understanding is being developed using a strategy 2 that:
• Tries to address the multiple perspectives of understanding identified in [Sanz et al., 2023]
(when relevant).
• Is driven by the needs of the domain —cognitive robotics— and the three testbeds; it is
applied in them and gets feedback from them.
In the remaining of the chapter we will address what shall be the content of such a theory and
what are the steps that implement this strategy.
2.2 On Theories
The Theory of Understanding (ToU) shall be a theory that is both scientific and operationalis-
able. As a scientific theory, the ToU shall be a well-substantiated explanation of some aspect
of the natural world that is based on a body of evidence, observations, and experiments. This
body of evidence comes from the cognitive operation of humans and also from the cognitive
operation of robots, esp. in the domains of the testbeds. The ”aspect of the natural world”
that we are interested in, is the phenomenon of ”understanding”.
We want the ToU to be a solid theory —a robust and well-established scientific explanation—
to serve as the framework upon which scientists could build their understanding of the phe-
nomenon of understanding and pile-up solid research results, and engineers use this under-
standing in its operationalisation as applied science in building better cognitive robots.
2.2.1 Characteristics of a Scientific Theory

This is a list of key components and characteristics of a scientific theory and to what extent the
ToU shall address them:
Empirical Basis: A scientific theory is grounded in empirical evidence. It should be supported

by a substantial body of observations, experiments, and data collected through system-
atic and repeatable methods. This empirical support is crucial in distinguishing a theory
from a mere hypothesis or conjecture. The source of evidence is deployed human cogni-
tion and the situations where understanding play a central role. Obviously we are dealing
with cognitively problematic situations [Norman, 1980] and hence, a proper organization
of evidence will be critical. Ontologies will play a central role in supporting this.
Consistency: A scientific theory must be internally consistent, meaning its various compo-
nents and principles should not contradict one another. It should provide a logical and
coherent framework for explaining observed phenomena. It shall also be consistent with
other accepted scientific theories. The formal methods of WP1 and the model-based
methods of WP4 will provide the necessary support to guarantee this consistence.
16-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

2.2. On Theories
Testability: Scientific theories are testable and falsifiable [Popper, 1959]. This means that they
can be subjected to experimentation and observation, and there should be clear criteria
that, if not met, would disprove the theory. The ability to test and potentially disprove
a theory is a fundamental aspect of the scientific method. The KPIs of our testbeds will
provide the necessary evidence.
Predictive Power: A strong scientific theory can make predictions about future observations
or experiments. These predictions should be based on the theory’s principles and should
be verifiable through empirical testing. The theory’s ability to make accurate predictions
lends further credibility to it. This is an essential aspect for a theory that is used as a
design asset in an engineering endeavour. Engineers will use cognitive patterns based on
this theory to guarantee future effectiveness in the robot systems under construction.
Scope and Explanatory Power: A scientific theory should have a broad scope, meaning it can
explain a wide range of related phenomena. The more phenomena it can explain, the
more powerful and influential the theory is considered. This is at the core of our ambition.
We expect to develop a theory that not only addresses robotic understanding but general
cognitive systems understanding.
Simplicity: Simplicity, often referred to as the principle of Occam’s razor, suggests that if two
or more theories explain the same phenomena equally well, the simpler one is generally
preferred. Simplicity makes theories more elegant and easier to work with. So far, there
is no real competing theory of understanding (i.e. in the terms stated here), nevertheless,
we will try to make the ToU simple using compact and effective abstractions. The use of
category theory is in part motivated by this search for simplicity.
Reproducibility: The results and observations that support a scientific theory should be repro-
ducible by other scientists using the same methods and conditions. This is a cornerstone
of the scientific method and ensures the reliability of the theory. This has always been
a problem in cognitive robotics and in cognitive systems in general. Replication of sys-
tems and experimental conditions is not easy. To this end, benchmarking has been used
in robotics to enhance this reproducibility. The RoboCup@Home used in WP8 specifically
addresses this problem. Concerning the robot system implementation, the model-based
methods of WP4 will provide the necessary means to achieve reproducibility.
Peer Review: Scientific theories are typically subjected to peer review, where experts in the
field assess the theory’s validity and the quality of the evidence supporting it. Peer review
is an important process for maintaining the rigor and reliability of scientific theories. The
Open Science approach of CORESENSE will specifically address this need. Besides formal
deliverables and publications we may produce specific materials for this need.
Consensus : Over time, a scientific theory gains acceptance and support from a significant
portion of the scientific community. While scientific consensus can evolve and change,
a well-established theory is generally widely accepted within the scientific community.
In the domain of psychology this is not easy. Psychological constructs suffer from the
”toothbrush” problem: no self-respecting psychologist wants to use anyone else’s. But in
rigorous science, we are sometimes forced to do so by the force of the facts. If our theory
is solid, well documented and effective, consensus will emerge. This is in fact, one of the
major challenges for the ToU and one of the major motivations for seeking a rigorous
approach.
Understandability : As quantum mechanics demonstrate, understandability is not a necessary
characteristic of scientific theory. However, we would like our theory to be understand-
able by a broad community of scientists and engineers (see Figure 1.1). This may require
CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05 | 17-49

from us the expression of the theory in different ways that can reach these people. This
is also important for the effectiveness of the work in WP5 and WP9-
It’s important to note that scientific theories are not absolute truths but are our best current
explanations based on the available evidence. They are subject to revision or even replacement
as new evidence and understanding emerge. The process of forming and refining scientific
theories is an ongoing and dynamic aspect of the scientific endeavor.
2.2.2 Operationalisation of a Scientific Theory

The term ”operationalizable theory” is not a standard concept in scientific terminology, but
appears to be a combination of two important concepts: ”theory” and ”operationalization.”
Theory: As mentioned in the previous response, a theory is a well-substantiated explanation of
some aspect of the natural world that is based on empirical evidence and can be used to make
predictions and understand phenomena. Theories provide a framework for understanding and
explaining the relationships between variables and how they influence various phenomena.
Operationalization: Operationalization is a process in research and social sciences where ab-
stract concepts or variables are defined and measured in a concrete and observable way. It
involves specifying how a theoretical concept will be measured or observed in practice. This
is a crucial step in turning abstract theories into testable hypotheses and conducting empirical
research.
It is possible that the term ”operationalizable theory” is used to refer to a theory that has been
sufficiently developed and defined so that its key concepts and variables can be operationalized
for empirical research. In this context, an operationalizable theory would be one that can be
translated into specific, measurable variables or constructs that researchers can work with to
conduct experiments, gather data, and test hypotheses.
In social sciences and some fields of natural sciences, this process of operationalization is essen-
tial for testing and refining theories through empirical research. Researchers need to make the
abstract and theoretical aspects of a theory concrete and measurable so that they can design
experiments or collect data to support or challenge the theory.
In summary, while ”operationalizable theory” is not a widely recognised term, it likely refers
to a theory that has been developed to a point where its key concepts can be translated into
observable and measurable variables for the purpose of empirical research and testing. This
operationalization is a crucial step in the scientific method when examining the applicability
and validity of a theory in real-world scenarios.
2.2.3 Essential Elements for a Theory of Understanding

The CoreSense ToU is a complex construct that consists of several key elements or components.
These elements shall work together to provide a comprehensive and well-substantiated view
of some aspect of the natural and artificial worlds. In some sense this view is explanatory —
e.g. when applied to humans or animals— in another sense this view is operational —as when
it enables the construction of artefacts.
In any case, there are some specific elements that a theory typically includes. Table 2.1 de-
scribes some of these elements and how we plan to approach them in the CoreSense ToU.
It’s important to note that the specific elements and structure of a theory may vary depending
on the field of science and the nature of the phenomenon being studied. This may imply some
specifics of the CoreSense ToU when applied to certain classes of systems. However, these
elements collectively contribute to the development of a robust and well-supported scientific
18-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

2.2. On Theories
theory.
Table 2.1: Essential Elements for a Theory of Understanding
Element Content of the theory The CoreSense ToU

Phenomenon A theory is typically developed The phenomenon addressed by
or Domain to explain a specific natural phe- the ToU is the phenomenon of un-
nomenon or a particular domain derstanding by a cognitive agent.
of inquiry. It defines the scope of The domain is the domain of cog-
what it seeks to explain. nitive systems, and not only cog-
nitive robots.
Concepts The theory contains a set of well- This document addresses in partic-
defined concepts that are central ular the definition of the core con-
to its explanation. These concepts cept for the ToU: ”understanding”.
provide the vocabulary and frame- Further version will address the
work for discussing and under- complete conceptual framework
standing the phenomenon. in collaboration with the develop-
ment of the ontologies in WP1.
Hypotheses The theory often generates One of the purposes of the CORE-
specific hypotheses, which are SENSE testbeds is the formulation
testable predictions or statements of specific hypothesis to validate
about how certain variables or fac- the theory.
tors are related within the defined
domain. These hypotheses guide
empirical research.
Laws or Princi- A theory may incorporate scien- The laws and principles will be
ples tific laws or fundamental princi- based on both the analysis of cog-
ples that describe relationships or nitive processes and the transla-
patterns observed within the phe- tion of general systemic properties
nomenon. These laws and prin- to the project context. These laws
ciples are typically derived from and principles will be formulated
empirical data and observations. in the different formal languages
of the project (esp. SysML and CT).
IN some cases they will be cap-
tured as patterns (e.g. in WP3).
Causal Rela- The theory specifies causal rela- System causality modeling sits
tionships tionships between the concepts at the very core of the modeling
and variables involved. It explains relation and is hence a core aspect
how and why certain factors in- of the theory.
fluence others within the phe-
nomenon being studied.
Explanatory A theory should have a high de- This explanatory power will man-
Power gree of explanatory power, mean- ifest at the engineering level,
ing it can account for a wide range but will be also manifest at the
of observations and data within runtime system level where the
its domain. It should provide a robot will be able to provide expla-
coherent framework for under- nations by introspection.
standing diverse aspects of the
phenomenon.
Predictive A strong theory can make accurate The testbeds in WP6, WP7 and
Power predictions about future observa- WP8 will help evaluate the predic-
tions or experiments. These pre- tive power of the theory concern-
dictions should be based on the ing the mission-specific factors of
theory’s principles and should be each testbed.
verifiable through empirical test-
ing.
CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05 | 19-49

Mathematical In some scientific theories, es- The formal approach of CORE-

Models pecially in the physical sciences, SENSE includes the use of math-
mathematical models may be ematical models using category
used to describe and predict the theory.
behaviour of the phenomenon.
These models are based on math-
ematical equations that express
relationships between variables.
Empirical Sup- A theory is grounded in empirical This is one of the main purposes of
port evidence. It should be supported the three testbeds.
by a substantial body of obser-
vations, experiments, and data
collected through systematic and
repeatable methods. Empirical
support is crucial in distinguishing
a theory from a mere hypothesis.
Refinements Scientific theories are subject to The ToU will evolve along the
and Revisions refinement and revision as new ev- project and will be exposed to the
idence and understanding emerge. community following Open Sci-
They should be adaptable to ac- ence practices.
commodate new data and insights.
Consistency A scientific theory must be inter- This is the central purpose of ex-
nally consistent, meaning its var- pressing the theory using formal
ious components and principles models. Consistency is specifically
should not contradict each other. addressed by the model-based
It should provide a logical and co- engineering approach fostered in
herent framework for explaining the project.
observed phenomena.
Peer Review Scientific theories are typically As said before, the theory will
subjected to peer review, where be exposed to the community
experts in the field assess the the- of peers following Open Science
ory’s validity and the quality of the practices.
evidence supporting it. Peer re-
view is an important process for
maintaining the rigor and reliabil-
ity of scientific theories.
2.3 Steps for a Theory of Understanding in CORESENSE

As said at the beginning of this report, the building of the ToU will be done in stages. This
deliverable is the first release of the theory.
As stated in the GA, the ToU will be re-released in final form at the end of the project (in M48).
We also expect to have interim versions of the theory during project execution. Interim re-
leases will be made available through the project website and the Open Science repositories.
Besides the complexity and need of scheduling personnel effort, the main reason for the ToU
being developed in stages is that its development proceeds along the project in a feedback
relation between the theory, assets, and application threads of the project (see Figure 2.1).
The testbeds inform the theory development and the theoretical snapshots will be translated
into software assets that will be used in the testbeds. AT the very end, the testbeds could be
used as demonstrators of the technology for exploitation purposes.
Another important reason for having several stages is that the theory will be specified in dif-
ferent formalisms. The initial version (this one) uses set theory. Planned future versions will
use a) more abstract maths (category theory [MacLane, 1998]), and b) the Systems Modeling
20-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

2.3. Steps for a Theory of Understanding in CORESENSE
Theory Development
Domain Theory Theory

Theory Initial Theory Interim Theory Final
Analysis Evolution Consolidation
Testbed Testbed Testbed Testbed Testbed Application

Analysis Implementation Use Implementation Evaluation Demonstrator
CS-IMG-036 Testbed Application
Figure 2.1: Theory development proceeds along the project in a feedback relation with testbed
development.
Language v2 (a language for model-based systems engineering [OMG, 2022]).
CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05 | 21-49

22-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

3 Towards a Theory of Understanding
”Reasoning is but reckoning” [Hobbes, 1651, Chapter 5]
”My experience as a scientist has taught me that the comfort

brought by a satisfying and well-worded definition is only short-
lived, because it is certain to need modification and qualification
as our experience and understanding increase; it is explanations
and descriptions that are needed” [Medawar, 1984, p. 154]
This chapter contains the first formal statement on the theory of understanding. The core seed
element of the ToU is a definition of what is understanding. This will be the central concept that
provides a ground to the theory.
This chapter also describes the expected content for a full-fledged theory and defines a set of
auxiliary elements that will be used in the definition of understanding. The chapter ends with
some considerations concerning the core definition.
3.1 Building blocks for a Definition of Understanding

Before presenting a definition of understanding, it is important to distinguish the main ele-
ments that constitute the definition: the object of understanding, how understanding is en-
abled, and the use (capabilities provided by) of understanding.
The object of understanding. We will use the term phenomenon as it is widely used by dif-
ferent authors (Boon, De Regt, Ylikoski, Thorisson,etc.). Since the word ”phenomenon” can
have different interpretations depending on who is using it, we will provide a definition of
phenomenon extending and generalising the one presented by Thorisson. A phenomenon is
defined as a set of elements and their relations. The phenomena can appear in:
• A closed system. A closed system is a system that is completely isolated from its environ-
ment. In this case, a phenomenon would be a part of the system (subsystem). In order
to fully understand a system, it would be necessary to understand all the elements and
relations that constitute that system, i.e., all the phenomena that can appear in the sys-
tem. Although in engineering systems are open, sometimes it is convenient to consider
them as closed.
• An open system. An open system is a system that has flows of information, energy, and/or
matter between the system and its environment and which adapts to the exchange. In
this case, a phenomenon would be constituted by elements of the system and the en-
vironment and their relations. This is the more generic assumption, and it covers cases
where there are different systems interacting among themselves and with the environ-
ment.
23
Chapter 3. Towards a Theory of Understanding
System definition: A system is a set of elements Vi that are interconnected.

S ⊂ V1 × ... × Vn , S ⊂ ×{Vi | i ∈ I}
Open system definition: an open system is a set of interrelated elements where

those elements can be partitioned into three classes, objects representing the influence
from the environment to the system (inputs) X = {X1 × ... × Xm }, objects representing
the influence from the system to the environment (outputs) Y = {Y1 × ... × Yn } and
system states Z = {Z1 × ... × Zn }. The open system is defined as: OS ⊂ X × Y × Z
Figure 3.1 shows a general open system (with interrelated elements) with several intercon-
nected subsystems and where every subsystem is also composed of interrelated elements.
Figure 3.1: A general open system, it is composed of different subsys-

tems (system of systems).
Phenomenon in a closed system S: ϕs = (v1 × ... × vn ) is an instance of the system. The

totality of all phenomena is the system itself.
Phenomenon in an open system OS (system and environment): ϕos = (x1 × ... ×

xm , y1 × ... × yn , z1 × ... × zn ) is an instance of the system and its relation with the
environment.
Instance: particular event, situation, occurrence of the system (and the environment).
Figure 3.2 illustrates the concept of phenomenon in a general open system.

Two very simple examples may help understand the essential nature of these definitions:
Example 1: Water tank. A simple example of a system can be a water tank with a pipe at an
inlet port and a pipe at an outlet port. The system includes as main object the tank that receives
water input (from the system environment) and generates water output to the environment.
The tank has properties such as base area and height. The volume of water in the tank can be
understood as a property of the tank (time-varying). The tank is managed by a human user who
has to make decisions on how to act on the inflow and outflow of water in order to maintain
24-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

3.1. Building blocks for a Definition of Understanding
Figure 3.2: A phenomenon (in red) as an instance of a system.
certain volumes of water. In this case, the subject of understanding is the human user, and the
phenomenon to be understood is how the volume of water changes with respect the water
inputs and outputs.
Example 2: Thermostat. The system includes the room where the temperature is measured,
the thermostat mechanism (with sensors, actuators and display for temperature adjustment by
the user) and the conditioning system for heating or cooling the room. In this case, the subject
of understanding can be the thermostat mechanism and the phenomenon to understand is the
change of the room temperature in relation to the on/off actions of the climate control system.
How understanding is enabled
According to the analysis of the state of the art, a subject is considered to understand if: 1) it can
interpret and structure the phenomena (relation between objects) [Boon, 2009], a theory ex-
ists [de Regt, 2009], 2) it grasps the relationships and has cognitive control [Hills, 2016],control
the phenomenon [Ylikoski, 2009], 3) it grasps structure [Zagzebski, 2008], 4) it grasps how the
various elements in a body of information are related to each other [Kvanvig, 2003], 5) it has
mental representations that encode the right kind of dependence relations manifest in the
system [Lombrozo and Wilkenfeld, 2019], 6) it has a set of models [Thórisson et al., 2016].
All of them share the need to have a representation of elements and their relations (a system
structure). This fits with what is generally accepted as the definition of a model, a model en-
codes and represents the elements and relations that are modelled. Thus, understanding is
enabled by the existence of a set of models of the phenomenon to be understood.
The use of understanding
It is widely accepted that understanding a phenomenon should enable several cognitive abili-
ties, among which are usually cited: reasoning about questions, making inferences (and coun-
terfactual inferences), making predictions, recognising consequences of a phenomenon, giv-
ing explanations, answering What If questions, anticipating behavior or understanding similar
phenomena.
All of them could be summarised in a way that mainly consists of making inferences (including
predictions and retrodictions), giving explanations, sharing knowledge with other agents, and un-
derstanding similar phenomena. We share this common vision, but in the case of explanation,
although it is a desirable property, we concur with recent opinions arguing that understanding
CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05 | 25-49

can be achieved without the ability to explain.

Among the most interesting abilities are those that enable anticipatory behavior and causal
reasoning:
• Inference: Drawing a conclusion about something by using information that you already
have about it.
• Prediction: A statement of what will happen in the future (or to utilise present informa-
tion to infer a future event or state of affairs).
• Retrodiction: To utilise present information to infer a past event or state of affairs.
These are just examples of model use. The features presented are essentially related to the
nature of the models. Models have a key property: they are representations of something that
enable several modes of understanding by means of a set of possible model exploitations. The
work in WP3 will specifically address the different classes of exploitations in the realisation of
cognitive functions.
3.2 A definition of understanding

Once the three elements that constitute the definition have been explained and defined, the
following definition for understanding is proposed:
A subject S understands a phenomenon ϕ if it has a set of models Mϕ of ϕ and those

models can be used to make sound inferences about the phenomenon.
Subject is the individual (human or machine) that performs the action of under-
standing. Examples: aerial robot, social robot, physicist, economist, sociologist.
Sound inferences: are those that follow the modeling relation, i.e., that implica-
tions in the model correspond to causality in the system (modelled). The conclusions
inferred from premises using the model correspond to object states of the modelled
system.
A variant under consideration addresses the mission-centric nature of purposeful understand-

ing for agents that shall decide on actions:
A subject S understands a phenomenon ϕ if the subject S is able to infer facts from

observations of that phenomenon (using a set of models Mϕ of ϕ) that help the subject
S judge the observed situation in order to assess possible decisions about what to do
in that situation.
However, this definition may seem too restrictive, specifically addressing teleological agents.
This variant may match better the needs of an agent as related with its value system. This will
be addressed in the future Theory of Awareness (ToA).
3.2.1 Degrees of understanding

Understanding is not a black or white thing. It is commonly agreed that understanding has
degrees. In fact, complete understanding of a system may not be possible. According to the
26-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

3.2. A definition of understanding
Figure 3.3: Left figure shows partial understanding of the phenomena as there is no model of
the interaction between them, the right figure shows a model covering also the interaction
definition, complete understanding would imply having a complete and perfect model of the
phenomenon that is being understood. In general, a model cannot be complete and perfect,
as that would be the modeled object itself. A model is a representation of the modeled system
that captures (part of) its structure through some encoding mechanisms. Thus, understanding
is always limited in breadth and depth. The degree of understanding is defined by the quality of
the models, that is, by their completeness and accuracy.
Figure 3.3 shows two phenomena that occur in a system (red and blue). On the left (scenario
1), the subject has a model of both phenomena and can understand both of them, but there
are some consequences that are not understandable, as the subject does not have a model of
how these two phenomena interact (purple relation). On the right (scenario 2), the subject has
a model comprising both phenomena and their interactions, so higher-level understanding can
be achieved in this case.
3.2.2 Some considerations

The base definition of understanding presented in Section 3.2 shall be further elaborated in fu-
ture versions of the theory. In particular, there are many other aspects to be taken into account
to make the ToU address all the elements of a theory that were shown in Table 2.1.
This is a list of some relevant considerations concerning the theory:
• The definition is general enough to cover particular types (or cases) of understanding.
It covers understanding when time is of importance (for example, understanding a sit-
uation) and when it is not (for example, understanding a structure, such as a layout or
a map). The definition covers mechanistic and functional understanding, as both can be
achieved having the right type of models. These particular types of understanding will be
developed in interim versions of this document in relation to the development of testbed
implementations.
• We take the following as a definition of “A subject understands a system”: A subject un-
derstands a system if the subject understands all the relevant phenomena that can occur
in the system for a specific scope under study (cf. the previous discussion on mission-
oriented understanding).
CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05 | 27-49

• The definition serves for real or imaginary systems as long as they are defined.
• Complete understanding of a real system is not possible in general for real world systems.
This would imply having a complete and perfect model of all the phenomena that can
happen in a system. In principle, a model cannot be complete and perfect (as this would
be the modeled system); however, this may be the case for formal abstract systems that
can be isomorphic to their models (e.g. we can understand conics geometry by modeling
it with algebraic equations).
• Answering questions about a system includes making retrodictions and predictions of the
phenomenon.
• As indicated previously, giving explanations is a desirable ability of a cognitive agent (it

can serve to communicate inferences and as proof of understanding), although it is not
necessary to achieve understanding. However, this property is usually present in most
cases of understanding (in agents capable of verbal reporting). Notice that a false model
can be used to give explanations (that may result nonsense) .
• Notice that the set of models includes any type of model (physical, mathematical, and
computational —if considered different from mathematical— because this is the core
of the hybrid nature of the CoreSense architecture) as long as they provide the abilities
above-mentioned. For example, a map is a physical model that allows one to answer
questions about which route is the shortest (making predictions) or how someone can be
at some point starting from another (retrodictions).
• Other features or properties are provided through understanding, like, for example, goal
achievement or making inferences about analogous phenomena. Although understand-
ing is necessary for these additional features, it is not sufficient. In the case of goal
achievement, meaning (in the sense presented by [Thórisson et al., 2016]) is also needed
(this will be the basis for what in this project is called situational awareness).
• In many (practical) cases, the phenomenon to be understood has to be perceived, and,

after understanding it, a meaning (implications) has to be derived, and, finally, usually an
action is proposed/made. This implies that understanding in most of the situations is not
an isolated activity, but one step more in a sequential process. Notice that in this case, if
the process (perception) prior to understanding does not happen or it is incorrect, it will
make that no understanding or bad understanding is produced. However, what is faulty
is the perception and not the understanding process itself.
• Understanding is enabled by a set of models. It happens that many times the concrete
model of the observed phenomenon alone enables only low-level understanding. In or-
der to achieve higher levels of understanding, this model has to be integrated with other
(existing) models, and, in this way, more consequences can be predicted (or more diag-
nosis can be made).
• The lack of a model implies that the phenomenon cannot be understood.
• A bad model implies that the phenomenon is either not understood or misunderstood.
• An observation with errors (which is taken as correct) prevents understanding or pro-

vokes misunderstanding (in the presence of a good model).
• It can happen that a subject has a model but the model has missing/uncertain data, in
that case it is possible for the subject to understand, but the inferences (predictions, etc.)
28-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

3.2. A definition of understanding
done will be uncertain, only qualitative, vague, or probabilistic predictions (different fu-
ture scenarios) can be done in that case.
• Maybe a model cannot be executed on time (lack of time or lack of capabilities of the
subject); in that case, the derivation of timely actions cannot be based on deep under-
standing, or it shall happen in a qualitative way. We may develop some form of scalable,
anytime understanding to address this issue.
• The degrees of understanding derived from the limitations of the models are not the only
limiting factors of understanding. The types of inferences applicable to the model also
determine different degrees of understanding. For example, an agent can be able to do
predictions but not retrodictions due to limitations of its cognitive engines.
CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05 | 29-49

30-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

4 Analysis
To understand the world the way humans do, agents will need
the ability to interact with it like we do - and this is exactly why
research in causal cognition should be tightly linked with efforts
to design robots that can learn and utilize causal information
about their environments. [Stocking et al., 2022]
In this chapter we provide an initial analysis of the initial theory of understanding presented in
this document in terms of both the project objectives and the concrete systems to be built:
• We analyse to what extent this theory offers some novelty as a theoretical contribution
to cognitive science at large.
• We analyse to what extent this theory covers the different perspectives identified in
[Sanz et al., 2023] both from the theoretical point of view and the testbeds point of view.
• We analyse to what extent this theory is seen as promising to be useful for robots to
perform their duties and for their engineers to build such robots. We also analyse to what
extent it is useful as a general theory of understanding in the wider domain of cognitive
science.
• Finally, we analyse the potential future development of the theory using other formalisms
suitable to attain the project objectives.
4.1 Understanding and Models

In essence, our theory of understanding is a theory of model-driven minds. All we think about
cognition orbits around the creation and use of models to drive actions of the autonomous agent
(see Figure 4.1).
This may sound quite trivial and old: there are plenty of controllers and AI algorithms that use
models. Examples abound: from model-based controllers for wheeled robots, to statistical
models for traffic congestion control, or to deep neural network models for recognising cats
in pictures.
What is new then? What is the novelty of this theory? The answer is simple: the bits and pieces
are not so novel; its systemic aspects are.
The models shall not only be actionable, they shall also be integrable and multiplely
actionable when integrated.
31
Chapter 4. Analysis
The affirmation that all cognition is based on models from the very bottom, provides a
conceptual and operational structure for their effective integration even when they are
heterogeneous. For example, you may have a robot able to see (having vision models) and
able to hear (having audio models) but the proper handling of perceptual multimodality re-
quires their deep integration. This can be in an algorithm-centric, ad-hoc manner, or can be
based on model integration based on a formal systemic background framework.
In essence, besides the common use of the term ”understanding” in human language, under-
standing is not just language, but deep systemics: 1) on the side of the object, 2) on the side
of the subject, and 3) on the whole, as a system-of-systems (see Figure 4.1).
Figure 4.1: While ”understanding” is commonly tied to language

processing the deep understanding that we need for our robots
is not just language, it is systemic: the cognitive agent creates an
abstract system in its head that matches some specific aspect of
the reality.
4.2 Coverage
The theory we are seeking shall be operational in the construction of better machines, but at
the same time shall provide a glimpse into a theory of understanding in general cognition.
4.2.1 Coverage of perspectives

Given the perspective of understanding as a multifaceted phenomenon (see the dictionary en-
try in [Sanz et al., 2023]) we can check to what extent the current theory addresses all these
possible meanings.
As said before, all these uses of ”understanding” will be of relevance in CORESENSE (to differ-
ent extents). Obviously, the perceptual, epistemic, and doxastic aspects are of central impor-
tance for a robot cognitive architecture, but other uses like trust and agreement will also be
critical in dependable systems.
This was the list:
1. VARIABLE NOUN - in the sense of perception

If you have an understanding of something, you know how it works or know what it
means.
32-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

4.2. Coverage
2. ADJECTIVE - in the sense of sympathy

If you are understanding towards someone, you are kind and forgiving.
3. UNCOUNTABLE NOUN - in the sense of empathy
If you show understanding, you show that you realize how someone feels or why they did
something, and are not hostile towards them.
4. UNCOUNTABLE NOUN - in the sense of trust
If there is understanding between people, they are friendly towards each other and trust
each other.
5. COUNTABLE NOUN - in the sense of agreement
An understanding is an informal agreement about something.
6. SINGULAR NOUN [oft NOUN that] - in the sense of belief
If you say that it is your understanding that something is the case, you mean that you
believe it to be the case because you have heard or read that it is.
Remember what the definition says:
A subject S understands a phenomenon ϕ if it has a set of models Mϕ of ϕ and those models

can be used to make sound inferences about the phenomenon.
Table 4.1: Coverage of perspectives of understanding
Perspective Coverage by definition and/or use

perception The definition of understanding essentially targets the process of generat-
ing metal representations —beliefs— from processing of the sensory infor-
mation flow —the perception— as targeted at the agent’s mission.
belief As said, this is in the central scope of the definition.
sympathy, The aspects of sympathy/empathy are related to interactions with other
empathy agents. While the definition is certainly neutral concerning this, it enables
the construction of models of the others that have the necessary properties
to support operational subsystems that power these capabilities. The phe-
nomena in this case are the behaviors of others, and these are the entities
that are properly understood.
trust The nature of the modeling relation that is behind the definition of under-
standing is what makes the system trustworthy in an intrinsic sense. It is
not just the case that the user trusts the robot because of the users’ psy-
chology, but that the robot is intrinsically trustable because its beliefs effec-
tively match the reality that surrounds the user and the robot.
agreement Understanding as an agreement about something between agents is en-
hanced by an explicit, externalisable formulation of what both agents un-
derstand about the object of the agreement.
4.2.2 Coverage of the Testbeds' Needs

Deliverable D1.1 [Sanz et al., 2023] made some considerations in relation with the importance
of understanding in the different project testbeds (see [Sanz et al., 2023] Section Using Under-
standing in Machines). In this section, we will see if our current theory is suitable to address the
needs of the testbeds.
CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05 | 33-49

Chapter 4. Analysis
Coverage of the Manufacturing Testbed Needs

See [Sanz et al., 2023] Section Understanding in the Manufacturing Testbed to get the contexts
for these examples.
UC-MT-1.3: Part Inspection and Product Iteration
Following the understanding definition from Section 3.2, we shall identify the following as-
pects:
1. System: The main system in this scenario is the inspection area equipped with a high-
definition camera and advanced inspection algorithms.
2. Phenomenon to understand: The primary phenomenon in this scenario is the status of a

specific part using the camera and situated in a environment and its implications on the
process.
3. Provided understanding: In this scenario, understanding focuses mainly on taking informed

decisions that can help optimize the inspection process with knowledge about the status
of the part and its transportation destination accordingly. In concrete, understanding
shall provide information concerning the following aspects:
• Characteristics of defects, such as their size.

• Determination of the material composition of a defective part.
The insights derived from the understanding process can be leveraged by human operators to
adjust manufacturing parameters effectively, reducing the likelihood of defects during reiter-
ation or when producing a new batch; and more straightforward to decide the next action to
take depending on the piece status.
UC-MT-1.5: Human Robot Coordination and Cooperation
Following the understanding definition from Section 3.2, we shall identify the following as-
pects:
1. System: The main system is the cobot which interacts with a human operator and the trays
and parts for assembly.
2. Phenomenon to understand: The phenomenona involved in this scenario is the human

robot-collaboration in the assembly process, which includes the human behavior in the
cell and the implications of the cobot motion on safety.
3. Provided understanding: In this scenario, understanding focuses mainly on how to reg-

ulate the cobot’s speed based on the tracked position of the human co-worker and its
intention to increase safe collaboration. In concrete, understanding shall provide infor-
mation information concerning the following aspects:
• Predictions regarding how this distance is likely to change based on both human
behavior and the cobot’s actions.
• The predictions above can later be used for dynamic adjustment of operation zones
in response to real-time conditions and safety requirements.
The insights derived from the understanding process should provide a comprehensive determi-
nation of the operation zones responsible for the cobot’s speed regulations. The propositions
derived from understanding may contribute to enhancing the efficacy of the assembly process
34-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

4.2. Coverage
while maintaining a high level of safety. It ensures that the cobot’s speed is synchronized with
the human’s position, speed, and intentions, thereby optimizing overall performance.
To attain such a comprehensive understanding, the system shall handle some knowledge on as-
pects such as the reliability of laser scanners and the functioning of 2D positioning algorithms,
which are essential for accurately estimating the distance between the human and the robot.
It also involves the development of a human intentional model, which considers factors such
as its speed and the current assembly status to determine whether the human co-worker in-
tends to monitor the process or retire a completed piece. Additionally, knowing the principles
of operation zones is crucial, as it determines when and for how long adjustments should be
made to ensure optimal operation.
UC-MT-2.1: Driving between workcells
During the execution of this scenario, the following questions emerge:
• How can we achieve precise localization in this scenario?

• What strategies can be employed to plan a safe path, given the unreliability of the static
map and the need to avoid restricted areas?
• How can we ensure that the robot’s movements remain predictable in these challenging
conditions while simultaneously avoiding dynamic obstacles?
Gaining a solid understanding of this process can greatly improve its performance. According
to the definition discussed in Section 3.2, we shall identify the following aspects for under-
standing:
1. System: The system in this scenario is the mobile robot with the navigation task. How-
ever, in this case, the complex environment in which the system is deployed is especially
relevant. It is an industrial plant where paths may be blocked or restricted, there may
be humans walking and moving forklifts carrying heavy pieces. Moreover, there may be
large parts added or removed from the surroundings.
2. Phenomenon to understand: The primary phenomenon in this scenario is the process of
localization in a complex environment with static objects changing its position, dynamic
obstacles, and restricted areas.
3. Provided understanding: In this scenario, our primary goal is to improve the navigation
process while prioritizing safety. Specifically, our understanding aims to provide valuable
insights into the following aspects:
• Enhance scan-based localization by incorporating semantic context information.
• Improve robot control to prevent erratic movements.
• Identify how features of dynamic obstacles can affect the system and respond ac-
cordingly, taking into account their nature, expected speed, and trajectory.
To attain this degree of understanding, the system shall have knowledge, at least to some ex-
tent, about the different types of obstacle and their expected behavior, particularly when they
involve humans walking or forklifts carrying heavy parts. Furthermore, it requires having infor-
mation about the workspace and models that can effectively align with the robot’s localization
and planning algorithms. Moreover, the robot controller should be capable of integrating all
these factors to ensure a smooth and predictable robot motion. Such capabilities can signif-
icantly enhance navigation tasks, particularly within the complex environment of an aircraft
assembly facility.
CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05 | 35-49

Chapter 4. Analysis
Coverage of the Drone Inspection Testbed Needs

See [Sanz et al., 2023] Section Understanding in the Inspection Testbed to get the contexts for
these examples.
UC-IT-1.1: Battery power consumption
1. System: The system in this example is the physical body of the aerial robot that includes
the different hardware components such as rotors, cameras, etc. In particular, the electric
power is provided by a battery which is equiped with sensors that measure periodically
its charge.
2. Subject of understanding. The subject of understanding is the aerial robot that performs
the inspection task.
3. Phenomenon to understand. In this example, the phenomenon to understand is the bat-

tery power consumption.
4. Provided understanding: As it is explained below, the benefits of understanding the de-

scribed phenomenon about the battery power consumption are related to improved safety
and explainability.
Checking periodically assumptions about the battery power consumption allows the robot to
continue its flight more safely. If any of these assumptions fail, an alarm is raised that can be
used by the robot to change its flight plan. For example, if the battery has discharged faster
than expected, but there is still enough charge left to complete part of the mission, the flight
plan can be modified to cover that part and return to the starting point before the battery is
completely discharged.
The lack of such a battery charge understanding mechanism would mean that the robot would
have to make drastic decisions to ensure safe flight. For example, the robot could have a reac-
tive mechanism based on a safety threshold that lands the robot immediately when the charge
drops below the threshold. A limitation of this solution is that it may cause the robot to land far
away from the return point where the human operator is located and, in addition, an accident
may occur due to a forced landing in an unsafe place.
As an additional benefit of using the battery understanding mechanism, the user can obtain
explanations in order to better understand the reasons behind the decisions made. With such
information the user will be able to determine that the vehicle’s battery needs to be recharged
or, on the contrary, it should be marked as faulty due to malfunctioning.
UC-IT-4.3: Spatial distribution of the components of the photovoltaic plant
1. System: The system in this example includes the components of photovoltaic plant (e.g.,
panels arranged in sets of lines, etc.), the aerial robots that perform the inspection and
the humans involved in the inspection task (e.g., operators in charge of robots).
2. Subject of understanding. The subject of understanding is a specific aerial robot that per-
forms the inspection task.
3. Phenomenon to understand. In this example, the phenomenon to understand is the spa-

tial distribution of the components of the photovoltaic plant and the motion properties
(localization, speed, etc.) of robots with respect of such components.
36-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

4.2. Coverage
4. Provided understanding. The availability of the understanding mechanism allows, for ex-
ample, mobile robots to act by planning motion actions to reach user-specified goals, en-
visioning different courses of action and choosing the best ones. As specific advantages,
the spatial understanding process presented in this example allows for greater location
accuracy due to the redundancy provided by multiple aerial vehicles (derived from col-
lective SLAM), and to communicate information to the user with more understandable
concepts (panels, panel lines, etc.), rather than simply through geographic coordinates
(derived from semantic SLAM).
UC-IT-4.2: Camera image defects
1. System: The system in this example includes the components of photovoltaic plant (e.g.,
panels arranged in sets of lines, etc.), the aerial robots that perform the inspection and
the humans involved in the inspection task (e.g., operators in charge of robots).
2. Subject of understanding. The subject of understanding is the aerial robot that performs
the inspection task.
3. Phenomenon to understand. In this example, the subject tries to understand camera image
defects.
4. Provided understanding. The lack of an automatic understanding mechanism for camera
image defects would require constant manual monitoring by the flight operator to ver-
ify the quality of the images. This alternative solution would significantly increase the
cognitive load of the human operator during flight monitoring and could also result in
the acquisition of faulty images due to human error due to attention lapses. As an addi-
tional benefit of using the image quality understanding mechanism, the user can obtain
more precise explanations of what is happening (e.g. types of camera defects) in order to
understand the reasons behind the decisions made. With such information the user can
determine whether the camera is operating correctly or whether it should be marked as
faulty for future calibration or replacement.
Coverage of the Social Robotics Testbed Needs

UC-ST-1.1: User approaches the robot and robot is aware
Following the understanding definition from Section 3.2, we identify the following aspects:
1. System: The user, the Tiago robot and the office space where the interaction takes place.
2. Phenomenon to understand: the phenomenon to understand is the intention of the hu-
man, the robot has a model of user movement, face, or whatever and uses it to predict
that the user wants to interact.
3. Provided understanding: in this scenario, a subject with understanding (the robot):
• Integrates multiple sensory input (readings about the human location and about
self-localization) in its model (lay out of the office and potential relevant locations
for humans in it) to reliable assign interaction intention to the human (which then
becomes a user for the robot).
• Executes appropriate actions, both external (motions scially compliant with facing
someone for interaction) and internal (predictions, belief’s updates), and monitors
its expected effect for the human to become aware of the robot awareness of the
human intention to interact
CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05 | 37-49

Chapter 4. Analysis
This scenario requires the subject of understanding to possess models of the office and ex-
pected human use of its areas, as well as a theory of mind and social rules for interaction, and
the ability to use them to predict and to evaluate the outcomes of its actions.
UC-ST-2.2: Implicit instruction
1. System: The user, the Tiago robot and the office space and objects where the interaction
takes place.
2. Phenomenon to understand: the primary phenomenon is the integration of the input from
the user in the internal model of the subject as a goal. There are two objects of under-
standing: the physical objects related to the task (object and locations), and the belief
about the goal of the user.
• integrates multi-modal information (speech and non-vernal behavior) into its inter-
nal model of the desires of the user.
• explains the result of that integration to the user using one or a combination of
channels (display, voice, non-verbal/pointing)
This scenario requires the robot to posses knowledge about the relationship between the ob-
ject and the locations in the environment, as well as the value of the object for the user, the
user plausible desires, and relationship between that knowledge and multi-modal information
(speech and non-vernal behavior).
UC-ST-3.2: Obstacles in my way
1. System: The user, the Tiago robot and the office space where the interaction takes place.
2. Phenomenon to understand: the primary phenomenon is the use of models to support the
planning and execution of actions to achieve the goal by generating meaning of interme-
diate states.
• Moves according to social rules.
• Decides appropriately whether to: a) move towards the humans blocking the way in
the hope they will leave way, b) request verbally from the humans to move aside, c)
request humans to reach the object and hand it over.
• Updates its plan based on the observed result of its actions in relation with its plan
and its belief about the humans desires.
In this scenario the robot shall have knowledge about human behavior and social rules, in ad-
dition to navigation skills.
4.3 Future development of the theory

As said in Section 2 the theory will continue its development until the end of the project. The
current definitions will change, others may be added or eliminated, or the theory will lead to
some other developments. This section defines some of the already stated future develop-
ments.
38-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

4.3. Future development of the theory
4.3.1 Other directions of development

Besides the elaboration of the fundamental concepts of the theory, there are other directions
that we will explore:
• Artificial General Intelligence (AGI). The central topics of CORESENSE directly touch on
some of the central aspects of the Artificial General Intelligence research programme. We
will monitor their contributions and see to what extent they are aligned with ours. How-
ever, the AGI programme is essentially focused on achieving human-level intelligence,
while we try to remove any anthropomorphic bias.
• Metacognition. The ”cognition are models” adage can be applied to cognition itself. The
agent can model its own cognitive processes giving a path to higher-order thought pro-
cesses and self-awareness. This is an essential property for resilient, adaptive systems.
• Uncertainty. The modelling relation shall be able to handle the inherent uncertainty
present in the world and injected into the perceptual process.
• Explore the spectrum of forms of model use. Inference, as the exploitation of the deep
relation of the model with the causality of the object modelled, is not the only form of
model exploitation. A deeper understanding is achieved if the model can be used in other
ways (e.g. shared, merged, compressed, verified, etc.). Most of these ”other” forms of ex-
ploitation will be more related to the machinery of understanding than to the modelling
relation itself.
• Investigate the lifecycle of models. These other forms of model use lead to a broader
consideration of the life cycle of the mental models used by an agent. How are the mod-
els created? When shall they be eliminated? This points to the classic problems of truth
maintenance and nonmonotonic reasoning in logic-based systems.
4.3.2 Reification as reusable assets

One of the objectives of the project is that the theory of understanding will not be limited to
a theoretical formulation but will be reified in the form of reusable assets.
These assets will address three of the main stages of the autonomous system lifecycle:
System conceptualization: The theory will be translated into WP1 ontologies to be used in
the conceptualization of new systems.
System design: The theory will be translated into SysMLv2 models to be used in the systems
engineering processes supported by the WP4 tooling.
System implementation: The theory will be translated into ROS 2 packages to be used in the
robotics community addressed by WP5.
4.3.3 Towards a CT Definition of Understanding

In CORESENSE we try to elaborate formal theories of understanding and awareness based on
formal modelling languages. One of these languages is abstract maths, and, in particular, Cat-
egory Theory (CT). We think that the formalisation of these definitions can be expressed using
CT to achieve better levels of formal characterisation (see Appendix A for an introduction to
the main concepts of CT).
CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05 | 39-49

Chapter 4. Analysis
For example, following the human-readable definitions in Section 3.2, we can start by finding
three main categories. The Category of Systems (S), the Category of Behaviors (B) and the
Category of Values (V ).
We define a partition of the category of systems, a smaller category S. This category includes
the objects and morphisms of interest for a particular aspect, stakeholder, or situation, as de-
fined in Section 3. A functor M , establishes a map between the observed system and the model
of it.
M : Si → MSi
The model MSi is the result of applying a model functor M to the system of interest category
Si . Note that models —i.e. the modeling relation— are seen as functors because they provide
a notion of sameness between two categories [Rupnow et al., 2023]. The MSi category can be
seen as a simplified version of the system, complete enough for S partition.
In the coming months we will explore this approach by trying to identify the proper categories
and functors for both general cognitive aspects —e.g. self-awareness mechanisms— and also
for the three testbeds. We will also evaluate the use of CT in terms of both rigour and effective
use of the formalism in both engineering and runtime (e.g. by direct mapping of categories into
ROS-friendly datatypes).
40-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

Glossary
AGI Artificial General Intelligence. 39
AI Artificial Intelligence. 15
category theory Category theory is a general theory of mathematical structures and their re-
lations. 17, 20
cognitive pattern A reusable partial design to endow a system with a concrete cognitive ca-
pability. 17
CT Category Theory. 39, 40
doxastic concepts or theories related to beliefs, including the study of belief systems, their
formation, justification, and evaluation. 32
epistemic concepts or theories related to knowledge, including the study of justification, use,
transfer and truth. 32
functor in category theory, a functor is a mapping between categories. 40
metacognition awareness of one’s thought processes and an understanding of the patterns

behind them. 39
robotics An engineering discipline that deals with the mechanical structures and control sys-
tems of machines that move and act in the world. 15
ToA Theory of Awareness. 26
ToU Theory of Understanding. 16, 18, 20, 23, 27
41
Glossary
42-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

Bibliography
[Awodey, 2010] Awodey, S. (2010). Category Theory. Oxford Logic Guides. Oxford: Oxford
University Press, 2 ed. edition.
[Barr and Wells, 1990] Barr, M. and Wells, C. (1990). Category Theory for Computing Science.
Prentice-Hall International series in Computer Science. Prentice Hall.
[Baumberger, 2014] Baumberger, C. (2014). Types of understanding: Their nature and their
relation to knowledge. Conceptus, 40.
[Boon, 2009] Boon, M. (2009). Understanding in the Engineering Sciences: Interpretive Struc-
tures. In de Regt, H. W., editor, Scientific Understanding: Philosophical Perspectives, pages
249–270. University of Pittsburg Press.
[CORESENSE Consortium and EC, 2022] CORESENSE Consortium and EC (2022). CORESENSE
Grant Agreement. Technical report, The CORESENSE Consortium and the European Com-
mission Horizon Europe Programme.
[Craik, 1943] Craik, K. J. W. (1943). The Nature of Explanation. Cambridge University Press.
[de Regt, 2009] de Regt, H. W. (2009). The Epistemic Value of Understanding. Philosophy of
Science, 76(5):585–597.
[Fong and Spivak, 2019] Fong, B. and Spivak, D. I. (2019). Seven Sketches in Compositionality: An
Invitation to Applied Category Theory. Cambridge University Press.
[Hills, 2016] Hills, A. (2016). Understanding why. Noûs, 50(4):661–688.
[Hobbes, 1651] Hobbes, T. (1651). Leviathan, or The Matter, Forme, and Power of a Common-
wealth Ecclesiasticall and Civill. James Thornton, Oxford.
[Kozen et al., 2006] Kozen, D., Kreitz, C., and Richter, E. (2006). Automating proofs in category
theory. In Furbach, U. and Shankar, N., editors, Automated Reasoning, pages 392–407, Berlin,
Heidelberg. Springer Berlin Heidelberg.
[Kvanvig, 2003] Kvanvig, J. L. (2003). The Value of Knowledge and the Pursuit of Understanding.
Cambridge Studies in Philosophy. Cambridge University Press.
[Lawvere and Schanuel, 1997] Lawvere, F. and Schanuel, S. (1997). Conceptual Mathematics: A
First Introduction to Categories. Cambridge University Press.
[Lloyd, 2021] Lloyd, K. A. (2021). Category theoretic foundations for systems science and en-
gineering. In Metcalf, G. S., Kijima, K., and Deguchi, H., editors, Handbook of Systems Sciences,
pages 1227–1249. Springer Singapore, Singapore.
43
Bibliography
[Lombrozo and Wilkenfeld, 2019] Lombrozo, T. and Wilkenfeld, D. (2019). Mechanistic versus
Functional Understanding. In Grimm, S. R., editor, Varieties of Understanding, chapter 11,
pages 209–230. Oxford University Press.
[MacLane, 1998] MacLane, S. (1998). Categories for the Working Mathematician. Graduate Texts
in Mathematics. Springer New York, NY.
[Mason, 2003] Mason, R. (2003). Understanding Understanding. {SUNY} Series in Philosophy.

State University of New York Press.
[Medawar, 1984] Medawar, P. B. (1984). Pluto’s Republic. Oxford University Press, Oxford.
[Norman, 1980] Norman, D. A. (1980). Twelve Issues for Cognitive Science. Cognitive Science,
4:1–32.
[OMG, 2022] OMG (2022). Omg systems modeling language (sysml®) version 2.0 beta 1. part
1: Language specification. OMG Draft Adopted Specification ptc/2023-06-02, Object Man-
agement Group.
[Popper, 1959] Popper, K. (1959). The Logic of Scientific Discovery. Basic Books.
[Rupnow et al., 2023] Rupnow, R., Randazzo, B., Johnson, E., and Sassman, P. (2023). Same-
ness in Mathematics: a Unifying and Dividing Concept. International Journal of Research in
Undergraduate Mathematics Education, 9(2):398–425.
[Sanz et al., 2023] Sanz, R., Rodriguez, M., Molina, M., Zamani, F., Aguado, E., Miguel Fernán-
dez, D. P. a. G., and Hernandez, C. (2023). State of the art and needs for a theory of under-
standing and awareness. Deliverable D1.1 CS-026, The CORESENSE Project.
[Spivak, 2014] Spivak, D. I. (2014). Category Theory for the Sciences. {MIT} Press.
[Stocking et al., 2022] Stocking, K. C., Gopnik, A., and Tomlin, C. (2022). From Robot Learning
To Robot Understanding: Leveraging Causal Graphical Models For Robotics. In Faust, A., Hsu,
D., and Neumann, G., editors, Proceedings of the 5th Conference on Robot Learning, volume
164 of Proceedings of Machine Learning Research, pages 1776–1781. PMLR.
[Thórisson et al., 2016] Thórisson, K. R., Kremelberg, D., Steunebrink, B. R., and Nivel, E. (2016).
About Understanding. In Steunebrink, B., Wang, P., and Goertzel, B., editors, Artificial Gen-
eral Intelligence. Proceedings of the 9th Conference on Artificial General Intelligence (AGI 2016),
pages 106–117.
[Wikipedia contributors, 2023] Wikipedia contributors (2023). Theory — Wikipedia, the free
encyclopedia. https://en.wikipedia.org/w/index.php?title=Theory&oldid=
1168082115. [Online; accessed 18-October-2023].
[Ylikoski, 2009] Ylikoski, P. (2009). The Illusion of Depth of Understanding in Science. In

de Regt, H. W., editor, Scientific Understanding: Philosophical Perspectives, pages 100–121.
University of Pittsburg Press.
[Zagzebski, 2008] Zagzebski, L. (2008). On Epistemology. Wadsworth Philosophical Topics.

Wadsworth Publishing.
[Zagzebski, 2019] Zagzebski, L. (2019). Toward a Theory of Understanding. In Grimm, S. R.,

editor, Varieties of Understanding, chapter 7, pages 123–137. Oxford University Press.
[Ziff, 1972] Ziff, P. (1972). Understanding understanding. Cornell University Press.
44-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

Appendices
Bibliography
46-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

A Category Theory for Roboticists
In this Appendix we provide an introduction to Category Theory (CT) for robotics. CT allows us
to be completely precise about otherwise informal concepts [Kozen et al., 2006].
Note that this annex is just a basic description for robot builders; for further reading, refer
to classics like [Lawvere and Schanuel, 1997], [MacLane, 1998], [Awodey, 2010] from a mathe-
matical perspective, [Barr and Wells, 1990] from a computing perspective, and [Spivak, 2014],
[Fong and Spivak, 2019] for a scientific application perspective with a strong mathematical fla-
vor.
A.1 Category
A category C, is a collection of elements with relations between them. It constitutes an ag-
gregation of objects with an imposed structure [Lloyd, 2021]. To specify a category, we need
three constituents:
• A collection of objects, Ob(C),
• a collection of morphisms f : X → Y , mapping between objects X, Y ∈ Ob(C), and
• a composition operation g ◦ f : X → Z composing morphisms f : X → Y and g : Y → Z .
Additionally, categories must satisfy the following conditions:
• identity: for every object X ∈ Ob(C), there exists an identity morphism idX : X → X .
• associativity: for any three morphisms, f : X → Y , g : Y → Z , h : Z → W , the following

expressions are equal: (h ◦ g) ◦ f = h ◦ (g ◦ f ) = h ◦ g ◦ f
• unitality: for any morphism f : X → Y , the composition with the identity morphisms at
each object does not affect the result, idX ◦ f = f and f ◦ idY = f .
A simple example is the Set category, in which objects are sets, morphisms are functions be-
tween sets, and a composition operation is the composition between its functions.
A.2 Functor
A functor F , is a map between two categories C, D. It assigns objects to objects and morphisms
to morphisms, preserving identities and composition properties. Functors preserve structures
when projecting one category inside another.
Diagram A.1 represents a functor F : C → D, mapping between a category C composed of
C0 , C1 ∈ Ob(C) with morphism f : C0 → C1 , and a category D composed of D0 , D1 , D2 ∈
47
Appendix A. Category Theory for Roboticists
Ob(D) with morphisms g1 : D0 → D1 and g2 : D1 → D2 . Functors map are the dashed blue
arrows.
D0
g1
C0 D1 (A.1)
f g2
C1 D2
Functors can map between the same types of category, such as Set → Set, or between differ-
ent categories, such as Set → Vect, between the category of sets and the category of vector
spaces.
A.3 Natural Transformation

A natural transformation α, is a structure-preserving mapping between functors. Functors project
images of a category inside another; whereas natural transformations shift the projection de-
fined by a functor F into the projection defined by a functor G. Diagram A.2 relates these three
concepts. There are two categories C, D and two different functors F, G : C → D. The two
functors are linked by the natural transformation α : F ⇒ G.
C α D (A.2)
To specify a natural transformation, we define a morphism αc : F (c) → G(c) for each object
c ∈ C, such that for every morphism f : c → d in C the composition rule αd ◦ F (f ) = G(f ) ◦ αc
holds. This condition is often expressed as the commutative diagram shown in Diagram A.3,
where the natural transformation morphisms are represented as dashed arrows. This means
that the projection of C in D through F can be transformed into projections through G. The
commutative condition implies that the order in which we apply the transformation does not
matter.
αc
F (c) G(c)
F (f ) G(f ) (A.3)
αd
F (d) G(d)
48-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

HORIZON
EUROPE
Title D1.3 Theory of Understanding

Subtitle An initial formal definition
Ricardo Sanz (UPM), Manuel Rodriguez (UPM), Martin Molina (UPM), and Esther
Author
Aguado (UPM)
Date 2023/11/05
Reference CS-050
Version 1.0 Final
URL http://www.coresense.eu/doc/CS-050.pdf
© 2022-2026 The CORESENSE Horizon Europe Project

coresense.eu

Coresense D1.3

Uploaded by

Copyright:

Available Formats

Coresense D1.3

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Coresense D1.3

Uploaded by

Copyright:

Available Formats

CS-050-v1.

Ricardo Sanz (UPM), Manuel Rodriguez (UPM), Martin

NO WARRANTY — Please read the fine print:

Title: D1.3 Theory of Understanding

This is CORESENSE Deliverable 1.3. It contains an initial attempt at a formalisation of the

Chapter 1 contains an introduction to the theme and the document.

Chapter 2 discusses the strategy for the elaboration of the theory.

Chapter 3 proposes the initial snapshot of the theory.

Chapter 4 performs an initial analysis of the theory.

Annex 1 presents an initial basic reference of category theory.

CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05 | 3-49

Version 0.1 | 2023/09/01

Version 0.2 | 2023/10/20

Version 0.3 | 2023/10/20

Version 1.0 | 2023/11/05

4-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

2 Strategy for building the theory 15

3 Towards a Theory of Understanding 23

A Category Theory for Roboticists 47

6-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

3.1 A general open system, it is composed of different subsystems (system of systems). 24

8-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

2.1 Essential Elements for a Theory of Understanding . . . . . . . . . . . . . . . . . . 19

4.1 Coverage of perspectives of understanding . . . . . . . . . . . . . . . . . . . . . . 33

10-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

”If the organism carries a ‘small-scale model’ of external reality

1.1 Towards Understanding

Figure 1.1: Cognitive science sits at the convergence point

12-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

orous methods from systems theory.

1.2 Context and Motivation

CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05 | 13-49

1.3 Document Structure and Content

Chapter 2: Describes the strategy to build the theory.

Annex: Provides a brief introduction to category theory.

14-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

A theory is a rational type of abstract thinking about a phe-

2.1 Understanding in CORESENSE

2.2.1 Characteristics of a Scientific Theory

Empirical Basis: A scientific theory is grounded in empirical evidence. It should be supported

16-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05 | 17-49

2.2.2 Operationalisation of a Scientific Theory

2.2.3 Essential Elements for a Theory of Understanding

18-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

Table 2.1: Essential Elements for a Theory of Understanding

Element Content of the theory The CoreSense ToU

CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05 | 19-49

Mathematical In some scientific theories, es- The formal approach of CORE-

2.3 Steps for a Theory of Understanding in CORESENSE

20-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

Domain Theory Theory

Testbed Testbed Testbed Testbed Testbed Application

CS-IMG-036 Testbed Application

Language v2 (a language for model-based systems engineering [OMG, 2022]).

CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05 | 21-49

22-49 | CORESENSE D1.3 | CS-050 | 1.0 | 2023/11/05

”Reasoning is but reckoning” [Hobbes, 1651, Chapter 5]

”My experience as a scientist has taught me that the comfort