Thérapie assistée par robot
Cats and dogs can enhance patient comfort levels, lower blood pressure, and positively influence other health factors Similarly, the use of robots could provide these benefits while alleviating doctors' concerns about potential drawbacks associated with live animals, such as the need for feeding and exercise Additionally, employing robots eliminates the risk of neglect or abandonment that can occur with pets.
Robots offer significant advantages in long-term therapy that requires intense and repetitive movement exercises They can perform the same tasks consistently and accurately without fatigue, while also recording and adjusting to patient responses Additionally, robots hold great potential for tele-rehabilitation, allowing a therapist to theoretically conduct multiple therapy sessions simultaneously or utilize a webcam for remote therapy.
Robots designed for rehabilitation can perform professional tasks in structured environments, such as handling paper in office settings or conducting tests in laboratories, as well as assist with daily activities in both structured and unstructured environments, including playing games, educational tasks, eating, and personal hygiene This contrasts with industrial robots, which typically operate in structured environments with predefined tasks and require skilled personnel with technical knowledge In rehabilitation robotics, the focus shifts towards service robotics, where humans and robots collaborate on the same tasks, necessitating a strong emphasis on safety and human-machine interfaces These interfaces cater to users with limited programming knowledge or those with physical challenges using specific programming devices Therefore, it is crucial to pay close attention to usability conditions, as the user plays an integral role in executing various tasks.
The development of robots involves creating devices designed to plan and execute specific tasks or a series of tasks The complexity of these tasks and the level of human interaction required can vary significantly among different types of robots However, most robots necessitate some degree of human involvement to perform tasks that are more than moderately complex Robots are utilized across a wide range of applications, from industry to entertainment and home maintenance A contemporary focus within robotics is rehabilitation robotics, as machines become more accessible and their ability to assist humans continues to grow.
The advantages of rehabilitation robotics are numerous, as they enhance the physical therapy paradigm by allowing therapists to focus on patient support while robotic exoskeletons facilitate movement These exoskeletons enable consistent training regimens, monitoring patient progress and adjusting effort levels accordingly, or providing recommendations to the supervising therapist Furthermore, companion robots can offer psychological comfort through advanced technology and artificial intelligence These benefits have led to the initiation of various research and commercial projects aimed at developing companion robots, such as the Probo robot from the Robotics and Multibody Mechanics research group at Vrije Universiteit Brussel, supported by the King Baudouin Foundation since 2006; the Pearl prototype from the research teams at Carnegie Mellon University and the University of Michigan, introduced in 2005; and the Paro robot developed by the Advanced Industrial Science and Technology (AIST) in Japan, unveiled in 2006.
Robots are an emerging technology that bridges the gap between the physical and social worlds of humans The future of rehabilitation robotics is promising, with various innovative technologies under research that are expected to deliver significantly improved outcomes As technology advances and costs decrease, rehabilitation robotics is set to become more popular and accessible to a wider audience.
The EmotiRob project, which is the focus of my internship, is situated in the field of Robot-Assisted Therapy (RAT) It began with an experiment using the Paro robot and is further supported by a research initiative funded by the National Research Agency (ANR).
Expérimentations avec Paro
Numerous experiments have been conducted in the field of Robot-Assisted Activities (RAA), particularly focusing on Robot-Assisted Therapy (RAT) In this context, T Shibata from the AIST laboratory in Tsukuba, Japan, developed a robotic seal named "Paro," which has shown promise in therapeutic applications.
A study conducted on elderly individuals in nursing homes and young children facing extended hospital stays revealed that robots provide significant benefits The findings indicate that the use of robots can enhance the quality of life for these vulnerable populations, highlighting their potential role in healthcare settings.
1 Agence National de Recherche un confort aux gens (même si l'évaluation du niveau du confort reste objectivement un problème très difficile)
In this context, the VALORIA laboratory conducted an experiment using Paro robots, provided by T Shibata, to explore whether reactions and interactions with robots vary based on cultural context Generally, the French exhibit more skepticism towards new technologies compared to the Japanese The experiment took place in two different locations: Kerpape, where Paro was utilized by therapists with children aged 6 to 12, some of whom were in wheelchairs, and IMA, where Paro was used with adolescents aged 12 to 14 who had behavioral disorders, including autism The findings demonstrate that Paro serves not only as a means of interaction between children and therapists but also provides a form of affection and attention to the participants.
Néanmoins, l’expérimentation a montré que beaucoup de problèmes doivent encore être résolus :
From a mechanical perspective, the Paro robot is too heavy for children with muscular disabilities To enhance its usability, the robot's autonomy needs to be increased, and improved handling features should be made available for children with limited strength.
• du point de vue d'interaction : les enfants veulent parler au robot et être compris, le robot devrait donc pouvoir exprimer plus d'émotions
Based on their findings, researchers at VALORIA have developed a method for robots to convey a wider range of emotional expressions without increasing the number of degrees of freedom A study conducted with over 1,600 participants online demonstrates that six degrees of freedom is sufficient for expressing basic emotions such as joy, sadness, fear, anger, and disgust.
VALORIA has proposed the EmotiRob project to the ANR, aiming to develop a new robot that retains all the fundamental features of Paro—such as sensors, softness, and comfort—while addressing two key issues.
• L’interaction langagière : ceci exprime que les humains pourront parler au robot qui est capable de comprendre ce qu’ils disent et de leur répondre en
The expression of emotion is aimed not only at conveying the robot's emotional state but also at establishing an internal model of emotion that provides the robot with a semblance of human-like feelings.
Project EmotiRob
Le projet EmotiRob [3] consiste à concevoir un robot compagnon pour enfants handicapés ou devant subir une longue hospitalisation
Comme abordé dans la partie précédente, nos deux expérimentations sur des enfants handicapés ont montré deux directions principales dans lesquelles il est intéressant de travailler
The first consideration involves mechanical issues: the robot needs to be lightweight, easy to handle, and more manageable than Paro, while also requiring a high level of autonomy.
The second wave of human-robot interaction highlights the psychological comfort that robots can provide, which is closely tied to the emotional bond a child forms with them This connection can be significantly enhanced if the companion robot possesses basic speech comprehension abilities and can express emotions in response.
Il convient donc de concilier les contraintes d’un robot léger et autonome à celles de capacités de compréhension et d’expression
As part of the project, a study was conducted on emotional expression, suggesting that six degrees of freedom are sufficient for a face to satisfactorily convey the six classic fundamental emotions Additionally, our laboratory focuses on research themes such as automatic speech understanding and disabilities.
The project aims to create a robot that retains the essential qualities of Paro, such as being a plush toy with a pleasant touch and equipped with sensors Specifically, we intend to develop a teddy bear-shaped robot that possesses the necessary capabilities for perception and natural language understanding, enabling it to form a formal representation of its user's emotional state Additionally, the project includes designing a model for the robot's emotional states and their evolution, ensuring that its reactions appear as natural as possible.
The EmotiRob plush robot stands out for its ability to respond to a child's behavior by simulating emotions through body movements, facial expressions, and simple sounds This innovative project lies at the intersection of robotics, human-machine communication, psychology, medicine, and vision, focusing on the detection, simulation, and production of emotions.
Il pose en particulier plusieurs problèmes scientifiques intéressants parmi lesquels :
• comprendre le langage d'un enfant qui parle à sa peluche et dans le même temps détecter son état émotionnel,
• doter le robot d'une capacité de réactions émotionnelles pertinentes,
• exprimer de manière intelligible des émotions par le robot,
• évaluer la qualité de l’interaction entre le robot et l’enfant qui revient à mesurer le niveau réconfort induit chez l’enfant
The impact of robots on children is a challenging issue due to the limited research available on this topic Consequently, this concern was underestimated during the development of the EmotiRob project.
Stage de Master
To assess the impact of robots on children, it is essential to understand their emotional state Since it is impractical to equip children with devices to measure blood pressure or heart rate, relevant information must be gathered through careful external observation of their emotional responses.
The issue of internships is now linked to the concept of emotion Specifically, this Master's internship focuses on defining and modeling emotions to create a framework that identifies observable items.
Dans le cadre du stage nous développons dans un premier temps un premier modèle d’émotion : GRACE La recherche a consisté d’abord en une étude bibliographique sur les émotions :
• A travers les travaux en psychologie et en philosophie sur les émotions chez l’humain (comme ceux d’Ortony, de Lazarus, de Scherer )
• A travers les travaux visant à construire un modèle calculatoire d’émotions (les modèles existants comme robot Kismet de Breazeal, EMA de Gratch et Marsella, FLAME d’El-Nars et al .),
To validate the relevance of the GRACE model, initial experiments have been conducted to simulate a very basic emotional process The goal of these experiments is to determine whether an observer, without prior knowledge of the project, can recognize the GRACE instance as embodying emotions If successful, this recognition would affirm the legitimacy of the model instance.
Structure du mémoire
This report outlines my experiences during my internship at the VALORIA laboratory, University of Bretagne-Sud, under the guidance of Professor Dominique Duhaut, as part of the EmotiRob project.
Le mémoire est divisé en cinq parties :
The first section introduces the internship topic, providing an overview of the field of Robot-Assisted Therapy It also includes information about the EmotiRob project, which is the context for the internship, and outlines the specific objectives of the internship.
The second section of the article presents a bibliographic study that includes definitions of emotion as proposed by philosophers and psychologists (§2.1), as well as recent computational models that simulate emotion for various applications (§2.2) This section ultimately concludes that there is no consensus on the definition of emotion.
In the third section, after discussing various reference models from the literature, we introduce our proposal: the GRACE model—Generic Robotic Architecture to Create Emotions We also provide justifications to demonstrate the general applicability of our definition of emotion in comparison to existing frameworks.
The implementation and validation of our model are discussed in the experimentation section (§4) During my internship, I was only able to program a reduced instance of the model, referred to as GRACE The validation approach involves testing the model's relevance in an ascending manner, starting from the simplest module and progressing to the more complex ones This reduced instance is designed to validate an initial segment that incorporates only three basic emotions: joy, fear, and sadness, along with two personality types.
The final section of the article presents the conclusions and discussions regarding the completed work It revisits the conducted studies while also addressing the remaining issues related to the topic.
The report concludes with an appendix (§6) and references (§7), which provide additional information about the program's data and links to the documents reviewed during my research.
To assess the quality of interaction between a robot and a child, it is essential to measure the level of comfort experienced by the child, which is closely linked to their emotional state However, defining what constitutes an emotion presents a challenge The question "What is an emotion?" has garnered significant attention from researchers seeking a universally accepted definition of human emotion.
Etudes sur la définition de l’émotion humaine
For a long time, human emotion has been the focus of extensive scientific research aimed at defining and understanding its components Initially, emotion was perceived as a mental concept, leading to thorough analyses by physiologists and psychologists Over time, studies have demonstrated that human activities are significantly influenced by emotional states, prompting the integration of emotional aspects into research on daily human activities such as communication, negotiation, learning, and commerce The increasing importance of computer applications in psychology, neurology, evolutionary biology, and even economics aims to enhance human-computer interaction Consequently, computer scientists are exploring new approaches to improve their programs, with some addressing the challenge of incorporating emotional elements This raises the question of what constitutes an emotion and how it can be deconstructed In our research, we have chosen to consult studies on the definition of emotion, particularly those from psychologists and physiologists, as their analyses may provide insights into the fundamental principles and functional mechanisms of emotion Below, we present a brief synthesis of the physiological, psychological, and philosophical work relevant to our subject.
Scientists have sought a relevant definition of emotion by examining its relationship with various existing concepts Philosophers have approached emotion from the perspective of mental topology, particularly focusing on its connections to bodily states, motivation, beliefs, and desires.
Research has shown that emotions possess an intentional structure, often linked to specific objects Additionally, emotions can be understood through their relationships with perception, cognition, daily actions, and self-knowledge Furthermore, emotions are also examined in the context of morality and rationality.
Tandis que ces sont des travaux importants et ne cessent à développer, il est difficile, voire impossible à avoir un accord entre les chercheurs pour une définition complète de l’émotion
La table ci-dessous (inspiré de [4]) présente une synthèse de quelques travaux philosophiques et psychologiques qui ont proposé des définitions avec des arguments intéressants :
Auteurs Définition Dans la relation avec
They argue that emotions are composed of basic emotions and are measured based on a limited number of finite dimensions, such as arousal level, intensity, pleasure or aversion, and intentions towards oneself or others.
Des émotions de base sont en général la joie, la tristesse, la peur, la colère, la surprise et le dégỏt
Ils mettent l’accent sur l’expression de l’émotion
Les recherches concentrent seulement sur l’expression des émotions de base (joie, tristesse, peur, colère, surprise, dégỏt)
Il a souligné à la fois l’intentionnalité et l’importance de la facture contextuelle sur la nature, l’arousal et l’expression des émotions
Contre la théorie qui dit que les émotions se composent principaleme nt des sentiments
Introduction de la notion d’appraisal impliquant que l’ộmotion entraợne l’évaluation
Approches psychologiqu e et de la biologie évolutionnair e
Il a défendu que l’émotion pourrait être vue comme l’état intentionnel
Il proposait aussi la notion de l’objet formel d’un état intentionnel
Contre la théorie qui dit que les émotions se composent principaleme nt des sentiments
Cet auteur a la même intention que celle de Bedfold
Il défendait l’idée que l’émotion relie fortement à la croyance et au désir
Défendre les théories des cognitivistes
(des philosophes qui suivent les théories cognitivistes)
Ils considèrent que les émotions impliquent les attitudes propositionnelles
Cognitions Les émotions dans ces théories cognitivistes sont caractérisées principalement en terme de leurs cognitions associées
Emotions are judgments defined by their temporal nature and evaluative content They may also be seen as strategic choices aimed at self-protection and enhancing self-esteem, drawing inspiration from Sartre's theory (1948).
Suivre les théories de cognitivistes
Selon lui, l’émotion commence à partir d’un changement physiologique anormal provenant d’une évaluation
Appuyer sur la relation entre l’émotion et l’évaluation
Les émotions sont émergées depuis la dynamique de l’interaction sociale
Dépendance de l’émotion sur son contexte
Inspiré depuis la théorie d’Errol Bedfold concernant la dépendance contextuelle de l’émotion
Les émotions sont les sentiments causés par des changements des conditions physiologiques concernant les fonctions autonomes et de moteur
A significant issue with this theory is its inability to adequately account for the differences between emotions The key distinctions among specific emotions are not physiological; rather, they are cognitive or based on other factors, such as the differences between anger and fear.
The author argues that emotion cannot be reduced to belief, desire, or a combination of both, but rather constitutes a logically and functionally distinct category of capabilities Emotions play a crucial role in the strategic choices involved in human rational deliberation.
La faỗon de percevoir le monde
Les émotions sont des réactions valencées aux événements, aux agents ou aux objets
(On revient à ce travail dans la section §3.1.1)
For her, emotion is a conscious mental process that significantly involves the body; it also plays a crucial role in shaping an individual's thoughts and actions, particularly in developing strategies for social interactions.
Considérer des émotions comme des processus internes
Ses travaux sont influencés par la théorie de jeu et les travaux des économistes
The appraisal process is essential and sufficient for emotional responses Additionally, it introduces the concept of coping, which enables individuals to select strategies to effectively address emerging challenges.
Klaus Scherer et ses collègues ont développé les théories d’appraisal à un modèle plus sophistiqué qui caractérise les émotions en terme de 18 dimensions (ou plus)
(On revient à ce travail dans la section §3.1.2)
Emotions are not merely isolated events; they encompass states and experiences that add depth, meaning, and motivation to an individual's actions.
Appuyer sur l’omniprésen ce de l’émotion
For him, emotions are primarily prototype concepts, and emotional modes represent a comprehensive exercise of all faculties, particularly in response to change, encompassing perception, intellectual processes, and feelings.
Emotions sont des modes subtils de l’actualisatio n mentale
Emotions can be understood as a complex process involving five key components: subjective feelings, cognitive evaluation, physical expression, action tendencies or desires, and neurological processes.
L’émotion sur le point de vue fonctionnel
Ces components peuvent être vus comme des émoteurs ayant un certain nombre de fonctions de base avec des algorithmes appropriés pour mener les processus physiologiques, expressifs, hormonaux et motivationnels
(On revient à ce travail dans la section §3.1.3)
Generally speaking, the definitions presented above vary significantly, often encompassing concepts such as context, intention, social interaction, strategic choices, belief, desire, motivation, internal processes, abnormal changes, experience, and bodily action Each definition is supported by its own arguments and rationale.
What exactly is an emotion? Is it possible to have a satisfactory definition that encompasses all previously discussed concepts or includes only essential elements? The answer is not straightforward In this work, we aim to strike a balance by formulating our own definition of emotion that aligns well with our EmotiRob project For more details, please refer to Section 3, which presents our proposed definition.
Etudes sur de récents modèles calculatoire des émotions
FLAME – Fuzzy Logic Adaptive Model of Emotion
The FLAME model, proposed by El-Nasr et al in 1999, is a computational framework for emotion modeling based on event evaluation It integrates learning components to enhance adaptability in emotional modeling and features an emotion filtering component that considers motivational states to address conflicting emotions Utilizing fuzzy logic rules, FLAME maps the impact of events on goals to emotional intensities Additionally, the FLAME agent employs a predefined reward value to assess the influence of user actions on the agent's goals.
En fait, le modèle FLAME contient trois composants : composant émotionnel, composant d’apprentissage et composant de prise de décision
The agent first perceives external events in its environment These perceptions are then relayed to both the learning component and the emotional component The emotional component processes these perceptions and also utilizes outputs from the learning component, including event-goal associations and desires, to generate emotional behavior.
This behavior is then relayed to the decision-making component to select an action The decision is influenced by the context, the agent's mood, emotional state, and emotional behavior; the chosen action will be initiated based on this decision.
Figure 2 - Composant émotionnelle du modèle FLAME
The emotional component is illustrated in the figure above, where boxes represent different processes within this component Information flows between these processes as depicted Initially, perceptions of the environment are evaluated through a two-step assessment process First, the experience model identifies which goals will be affected by an event and assesses the event's impact on these goals Second, mapping rules determine the desirability of the event based on the impact calculated in the first step and the significance of the associated goals.
The desirability measure, once calculated, undergoes an evaluation process to assess the emotional state change of the agent This triggers an emotion or a blend of emotions based on the desirability measure, which is then refined to create a coherent emotional state Subsequently, this emotional state is used in the behavior selection process, where criteria such as the current situation, the agent's mood, and emotional state are considered Ultimately, the emotional state may be diminished and fed back into the system for future iterations.
ParleE – Adaptive Plan Based Event Appraisal Model of Emotions 23
Introduced in 2002 by Bui et al., ParleE is a flexible and adaptive quantitative model of emotions designed for conversational agents in a multi-agent environment with multimodal communication capabilities It evaluates events based on learning and a probabilistic planning algorithm, while also modeling personality and motivational states to influence how an agent experiences emotions.
At the onset of an event, an Emotion Impulse Vector (EIV) is generated by assessing the event according to the rules proposed by Ortony et al., which are based on the agent's goals, plans, and standards The EIV encapsulates the impact of the event on emotions and is subsequently utilized to update the Emotion State Vector (ESV), which represents the intensity of those emotions.
Le Emotion Appraisal Component prend l’événement, la personnalité, le plan, et les modèles des autres agents comme entrée pour produire l’ESV en sortie Le
Planner produit un plan pour acquérir le but de l’agent Il calcule aussi la probabilité d’acquérir ce but Cette probabilité est utilisée par l’Emotion Appraisal
The Emotion Component calculates the Emotional Impact Value (EIV) and motivational states to produce an updated vector of emotions It collaborates with the Emotion Decay component, which assesses how emotions diminish over time, while also considering the influence of personality on emotional changes Additionally, the Model of Other Agents is employed to generate Desire-other emotions, reflecting the aspirations related to the fortunes of other agents.
La personnalité utilisée dans ce modèle est le modèle de personnalité de Rousseau
The model categorizes various processes an agent can perform, including perception, reasoning, learning, decision-making, action, interaction, indication, and sensation, all influenced by emotion However, it lacks clarity on how emotions specifically impact the planning process Additionally, the components of other agents' models appear to limit the flexibility of this model, contrary to the authors' assumptions.
Kismet - robot avec les émotions artificielles
This model aims to create an interaction between a robot, Kismet, and a human, reflecting the relationship between parents and their children during early communication Proposed by Cynthia Breazeal in 2002, this emotional model is grounded in agent-based architecture, where various system components operate in parallel and mutually influence each other.
Figure 4 - Architecture du modèle d'émotions de Kismet
This model was tested with five basic emotions: anger, disgust, fear, sadness, and joy, along with three additional emotions: surprise, interest, and excitement While the model has influenced the relationship between parents and children, it does not account for personality traits.
Greta – La dynamique de l’état affective dans un agent
In 2003, C Pelachaud and I Poggi introduced the Greta model, aimed at creating a human-computer interface through an animated conversational agent This model consists of two closely related components that work in tandem to enhance interaction.
This article discusses a dynamic mechanism for updating the representation of an agent's mind, which includes both long-term and short-term components such as personality and emotions It explores how these elements are triggered, deteriorate over time, and how the agent chooses to either conceal or express them Additionally, it examines the selection of bodily media for manifestation, such as gaze, voice, and facial expressions The formalism employed for this purpose involves dynamic belief networks with weighted goals.
The cognitive state of an agent is reflected in facial expressions that utilize various channels, such as fixed gaze direction, eyebrow shape, primary direction, and head movement This process involves resolving conflicting situations to determine the final expression, as different signals can overlap, mask one another, or be intensified or diminished.
Figure 5 - Réseaux dynamiques de croyance du modèle Greta
Although the model incorporates the concept of personality, this notion is not clearly defined Furthermore, the specific relationship between personality and emotion, as well as the influence of emotion on Greta's mindset, remains inadequately identified.
EMA – EMotion and Adaptation
In their research, J Gratch and S Marcella developed a comprehensive computational model of human emotion called EMA (Emotion and Adaptation) in 2004 This model aims to address two key factors: cognitive responses and behavioral responses, particularly focusing on coping mechanisms.
The current mental state of EMA is represented by a complex mental structure known as Causal Interpretation, designed to unify all the needs of an emotional agent into a simple architecture This causal interpretation consists of three interlinked components: the causal history (the past), the current world (the present), and the task network (the future) It takes into account various factors influencing emotions, such as plans, beliefs, desires, intentions, probabilities, and the utilities of events.
En modélisant EMA, les auteurs ont basé sur la théorie d'Ortony et al au sujet de l'évaluation d'événement, de la théorie de Lazarus au sujet d'évaluation et de coping
The model has been implemented to create an application where users can interact with virtual humans through natural language in high-pressure social settings However, the authors did not incorporate personality traits into the EMA model.
GALAAD – GRAAL Affective and Logical Agent for
Introduced in 2005 by Carole Adam and colleagues, GALAAD is an emotional conversational agent built on BDI (Belief-Desire-Intention) architecture, dialogue games, and speech acts Grounded in OCC theory of appraisal, GALAAD is capable of generating emotions that affect dialogue game rules and initiate a coping process as defined by Lazarus This coping strategy aims to maintain the agent's balance by reducing the intensity of negative or sensitive emotions that could adversely impact its behavior.
This model serves as an attempt to incorporate genuine evaluation and coping mechanisms into the architecture of conversational agents within dialogue games However, it overlooks the significance of personality and motivational states in emotional reasoning.
Another model developed by Carole Adam is the PLEIAD model, which appears to be an alternative version of GALAAD In this model, the author focuses on updating the agent's knowledge base by introducing a logical module for test shooting and activation The perceived stimulus is described by the user, detailing its name and effects; however, the expression of emotions through facial and bodily animation is not manipulated at all Like GALAAD, PLEIAD does not incorporate personality into its BDI agents.
Conclusion
Currently, there is no consensus on the definition or modeling of emotion Psychological research has explored various aspects of emotion, providing a more comprehensive understanding In contrast, computer scientists have proposed differing views on emotion This disparity motivates the ambition to propose a model of emotion that is sufficiently general to incorporate all essential elements identified in psychological studies while also encompassing previous computer science work.
In seeking a comprehensive model of emotions, it is essential to rely on credible psychological theories Notably, the works of Ortony et al., Lazarus, and Scherer stand out for their significant contributions to our understanding of emotional elements These studies have been analyzed and incorporated into computational emotion models discussed previously A novel aspect of our approach is the integration of personality into the emotional model, as personality plays a crucial role in shaping human emotional behaviors Indeed, personality is undeniable in our lives; it characterizes individuals and influences how we perceive, think, act, and adapt to daily life.
In this section, we aim to provide a detailed overview of the three psychological theories we reference, along with the MBTI personality standard we intend to incorporate into our emotional model Subsequently, we will present the proposed emotional model and justify its foundations based on the referenced theories.
Théories de références
Theory of Ortony & al (1988)
According to Ortony, Clore, and Collins, emotions are valenced reactions to events, agents, or objects, which are assessed based on an individual's goals, norms, and attitudes.
Figure 8 - Typologie de la théorie d'Ortony et al
The positive aspect of this theory lies in its close alignment with a computational approach It serves as a foundational model for understanding emotions due to its generic evaluation criteria However, a significant drawback of this framework is its failure to define the intensity of the resulting emotions.
Theory of Lazarus (1991)
According to Lazarus, individuals stabilize their relationship with the environment through two key processes: cognitive appraisal and coping Cognitive appraisal is defined as an adaptive process that helps maintain or alter the relationship between an individual’s beliefs and goals and the constraints of the world around them Lazarus identifies two types of appraisal: primary appraisal, which assesses the significance of an event, and secondary appraisal, which evaluates the potential responses to that event When a situation is perceived as stressful, individuals must adapt, highlighting the importance of coping strategies in managing stress.
• L’aspect de Problème-focalisé essayera de résoudre le problème (approche classique), mais pourra également nier le problème pour réduire au minimum l'effet
Emotion-focused coping strategies differ from denial by addressing the emotional response to a problem rather than the problem itself The focus shifts from the issue at hand to its physical consequences, emphasizing the importance of managing emotional reactions.
Theory of Scherer
Pour Scherer [15] cinq sous-ensembles fonctionnellement définis sont impliqués dans des processus émotionnels :
• Un sous-ensemble de traitement de l'information évalue le stimulus par la perception, la mémoire, la prévision et l'évaluation d'information disponible
• Un sous-ensemble de support ajuste l'état interne par la commande des états de neuroendocrine, somatiques et autonomes
• Un principal sous-ensemble projette, prépare des actions et les choisit entre les motifs concurrentiels
• Un sous-ensemble temporaire gère le moteur d’expression et le comportement visible
• Un sous-ensemble de moniteur finalement contrôle l'attention qui est assignée aux états actuels et passe la rétroaction résultante aux autres sous- ensembles
Scherer focuses on the subset of information processing known as stimulus evaluation checks (SEC) His theory posits that these evaluations significantly influence other subsets, leading to further changes in the processing of information.
Scherer propose cinq SECs substantielles, dont quatre entre eux possèdent d'autres subchecks
• Le contrôle de nouveauté décide si les stimuli externes ou internes ont changé ; ses subchecks sont précipitation, confiance et prévisibilité
• Le contrôle intrinsèque d'agrément indique si l'attraction est plaisante ou désagréable et cause des tendances appropriées d'approximation ou d'action d'éviter
The significance control of a goal determines whether an event supports or hinders an individual's objectives Its subchecks include goal relevance, outcome probability, expectation, supportive nature, and urgency.
Potential coping control assesses an individual's perception of their ability to manage events in their life This concept encompasses several sub-factors, including agency, motivation, control, power, and adaptability Understanding these elements can enhance one's ability to navigate challenges effectively.
• Le contrôle de compatibilité compare finalement l'événement aux normes internes et externes ; ses subchecks sont extériorité et intériorité
Chaque émotion peut, selon Scherer, ainsi être clairement déterminée par une combinaison des SECs et des subchecks Une table appropriée avec de telles attributions peut être trouvée dans [Scherer, 1988].
MBTI of Meyers-Brigg and Meyers
The Myers-Briggs Type Indicator (MBTI) was created to make Carl Jung's theory of psychological types accessible and practical for everyday life This individual assessment tool helps identify personal strengths and personality preferences As of 1991, over 10,000 people in the United States were using the MBTI daily.
Personality type is assessed by answering four key questions: where you focus your attention, how you process information, how you make decisions, and how you interact with the external world The evaluation process is outlined below.
Figure 9 - Standard de personnalité MBTI (1985)
Energy : Là ó vous focalisez votre attention : Préférez-vous se concentrer sur le monde externe ou sur votre propre monde intérieur ? Ceci s'appelle Extraversion
Extraversion (E) and Introversion (I) represent two distinct personality preferences Individuals who favor extraversion typically express themselves openly and engage readily with the external world In contrast, those who lean towards introversion are more inclined to reflect on their inner thoughts and impressions, focusing on their internal experiences.
When processing information, individuals may lean towards either Sensing (S) or Intuition (N) Those who favor Sensing concentrate on concrete details and what their five senses reveal in the present moment In contrast, individuals who prefer Intuition utilize their imagination to explore new possibilities and perspectives, focusing more on future potential rather than immediate realities.
When making decisions, do you prioritize logic and consistency, or do you focus on people and specific circumstances? This distinction is known as Thinking (T) versus Feeling (F) Individuals who favor Thinking typically base their decisions on objective analysis and logical reasoning, while those who lean towards Feeling often consider values and the well-being of others in their decision-making process.
In navigating the external world, individuals typically fall into two categories: Judging (J) or Perceiving (P) Those who identify as Judging prefer to make decisions and plan their lives, leading to a structured and organized approach Conversely, Perceiving types embrace spontaneity and flexibility, opting to remain open to new information and experiences This distinction highlights how different personality types interact with their environment, influencing their lifestyle choices and decision-making processes.
When you determine your preferences in various categories, you reveal your unique personality type, which can be represented by a four-letter code The identification and description of 16 distinct personality types arise from the interactions within these categories.
In our work, the MBTI model stands out as more intriguing than others because it establishes a flexible relationship between emotions and personality This adaptability allows us to create an emotional framework that can seamlessly integrate the MBTI, aligning with our goal of developing a comprehensive model of emotion that incorporates personality traits.
Modèle proposé : Generic Robotic Architecture to Create Emotions 36
Comme il n’y a pas de consensus sur la définition de l’émotion, nous voulons d’abord fixer une définition Il sera donc ensuite possible de proposer un modèle générique
Warning: We want to clarify that we use certain English terms to denote concepts These words are commonly used in everyday life to signify broader notions that may differ significantly from their meanings here We have chosen to use these terms to simplify the presentation.
Une émotion est un processus qui caractérise la réponse du corps humain à un événement
An event is the human body's perception of a change, or the lack thereof, in the environment, as well as a change, or absence of change, occurring internally within the human body.
The human body's response involves one or more internal changes that are undetectable to an external observer, as well as observable expressions or postures that indicate these changes, or potentially a lack of any noticeable alteration.
Basé sur cette définition nous proposons un modèle d’émotions de sept composants comme décrit dans la figure suivante :
In this model, sensation serves as the starting point, triggered by an event that may or may not exist, resulting in a physiological change within the body This sensation is processed through two parallel levels.
D’abord, la perception physiologique transformera ce signal initial directement en réponse du corps (battement du cœur, pression sanguine, etc.) et alarmera le niveau de comportement
La perception cognitive transformera ce signal de sensation en information cognitive sur la situation de l’environnement qui sera traitée par le comportement
Le comportement calculera la réponse à l’information venant du niveau des perceptions en se basant sur Internal cognitive state Cette réponse serait envoyée au
Body ó la réaction physique aura lieu
A noter que dans cette définition, le point de départ d’une émotion est un événement
Le processus de réaction à un événement prend quelque temps et ce temps n’est pas constant Alors, il est possible d’avoir une réponse à un événement suivant avant à celui antérieur
Defining this part is quite challenging, as it is where an emotion begins Following Scherer’s theory, the evaluation of an event constantly scans both the body’s environment and its internal state At any given moment, a change is detected, leading to the emergence of a sensation According to Ortony et al., this sensation can arise from an event, an agent's action, or a characteristic of an object This sensation is then processed by two interpretative modules—physiological and cognitive—each capable of inhibiting the input if its intensity is deemed too low.
The Sensation module can be viewed as a representation of the five human senses: sight, touch, hearing, smell, and taste It functions as a system of receptors that capture and translate various forms of energy (stimuli) for analysis, enabling perception and detection of changes in the environment The concept of senses encompasses two distinct aspects: immediate (instinctive) communication and mediated (rational) communication Moreover, senses are not merely transducers measuring parameters; they are instruments of perception, connecting the organism to the external world and facilitating the emergence of sensations.
The transformation from sensory activation to the sensations experienced in the body likely varies among individuals, influenced by factors such as sensitivity In the context of robotics, these senses are typically represented by sensors.
En répondant à une sensation, une sortie au niveau du corps pourrait être immédiate
Perception can be understood as an instantaneous event, reflecting either an external occurrence or an internal transformation triggered by an external stimulus For instance, a loud sound can overwhelm the auditory perception system, leading to a reflexive response in the individual This reflex arc consists of the transmission of sensory information to a nerve center, typically in the spinal cord, and the subsequent relay of a motor response from that center to the effectors in the body.
In our model, the physiological interpretation analyzes information from the Sensation module to calculate an immediate response, which is then provided to the Body module This response can also be transformed into information that serves as input for the Behavior module.
Cognitive interpretation acts as a filter for sensations, transforming them semantically and attaching meaning influenced by an individual's mood This mood amplifies specific characteristics of the sensation, forming the first part of the interpretation process The second part involves factors such as beliefs, novelty, and the alignment between personal standards and goals.
This type of interpretation is time-consuming and involves a cognitive process influenced by various factors Primarily, the internal cognitive state of the individual, characterized by mood and emotions, plays a crucial role Mood is shaped by personal history, fatigue, and concentration, affecting how events are perceived based on their valence Emotions and feelings represent a more complex internal state, with many aspects of emotions applicable to feelings Ultimately, feelings can disrupt interpretation by assigning positive or negative values to events.
Behavior calculates the emotional and behavioral response the body must provide to a perception, incorporating classic robotics elements such as planning, learning, and evolutionary methods Unlike traditional architectures, this approach establishes a dependency on the internal cognitive state Consequently, the reaction to a perception can vary, allowing for the distinction of two levels of responses.
• Le premier niveau étant donné classique est que quand on apprend que la réponse n’est pas adaptable aux entrées, on va calculer une autre réponse à la situation
The second level reflects the impact of cognitive states on emotional responses An aggressive or joyful cognitive state elicits different reactions compared to a calm or depressed one.
L’Etat cognitif interne (Internal cognitive state) est le lieu ó deux déclarations seront activées : sentiments (Feelings) et l’humeur (Mood)
Sentiments operate at a meta-level where cognitive perception, behavior, and bodily actions are analyzed This level examines the overall situation, encompassing feelings about previously experienced scenarios, a sense of moving in a positive direction, discomfort due to unmet expectations, or joy when circumstances feel manageable and under control.
Mood represents the overall storage of past emotions and significantly influences cognitive perception It encompasses various aspects, including behavioral responses (such as fight, flight, assistance, and affection), mental states (like motivation, interests, extraversion, and introversion), and physical conditions (such as fatigue and anxiety).
Intégration des caractéristiques du processus émotionnel dans GRACE
Event appraisal
The theory proposed by Ortony et al outlines the process of sensation and its internal processing When an event occurs, whether due to external changes or internal shifts, it is captured by the Sensation module and subsequently interpreted for further transformation This internal analysis of the event considers the relationship between current goals and interests, as discussed in Ortony et al.'s evaluation of events.
Coping
In our model, we focus on addressing coping strategies within internal analysis Generally, coping strategies impact the rules governing information processing, from interpretation levels and internal cognitive states to behavioral decision-making The outcome of this influence manifests as the behaviors or actions expressed by the body.
Cinq sous-systèmes de l’émotion
Deuxièmement, pour comparer notre modèle à la théorie de Scherer ci-dessus, nous avons :
• Le premier sous-ensemble (sous-ensemble de traitement de l'information) est le processus évaluant l'information du module de la sensation au niveau d'interprétation puis finalement au module de comportement
• Le sous-ensemble de support sera mis en application dans le module de l'interprétation psychologique et du corps pour ajuster l'état interne
• Le principal sous-ensemble sera mis en application dans le module de comportement
• Le sous-ensemble temporaire est intégré dans le module du corps
• Le sous-ensemble de moniteur sera développé dans le module des états cognitifs internes.
Personnalité de type MBTI dans GRACE
Le MBTI propose quatre catégories pour établir la personnalité Notre modèle couvre complètement ces quatre catégories
The first aspect is the Energy expressed through Extraversion (E) or Introversion (I) In the generic model, this specific mechanism is integrated within the Mood and Behavior modules Extraversion and Introversion serve as filters in cognitive perception, with Extraversion focusing on the immediate meaning of a sensation, while Introversion seeks a deeper interpretation Additionally, within the Behavior module, Extraversion leads to quick responses, whereas Introversion emphasizes finding the most appropriate reaction through introspection.
La deuxième catégorie du MBTI est complètement couverte par l'architecture générique La Sensation est construite avec les deux modules d’interprétation et l'iNtuition par le module d'Intuition
The third category is Decisions: Thinking (T) or Feeling (F) We explore these two approaches by coding the behavior module of the generic model If we create a decision tree with a detailed exploration of the solution, we will be in Thinking mode; however, if we code the behavior using a rule-based approach, we will operate in Feeling mode.
In conclusion, the MBTI lifestyle is categorized into Judging (J) and Perceiving (P), which influences how individuals interpret their experiences This classification reflects a preference for sensation, with those who identify as Perceivers often finding direct sensations related to others more engaging.
GRACE par rapport aux modèles informatiques récents
To achieve the goal of developing a comprehensive model that encompasses all existing models, we will compare our proposed model with those examined in the literature review This comparison aims to assess the validity of the model's generic nature.
On commence avec le modèle FLAME de Bui et al :
The blue arrows illustrate how the emotional component processes of FLAME will be implemented in GRACE In fact, the GRACE model is more complex than FLAME, as it delves deeper into internal cognitive states and the processing of incoming events Furthermore, GRACE will incorporate the learning and decision-making aspects that replace the two other components of FLAME, meaning our model effectively encompasses FLAME.
Nous allons ensuite regarder la ressemblance entre GRACE et ParleE de de Bui et al proposé en 2002
The figure below demonstrates that GRACE effectively simulates the processes of ParleE In fact, GRACE incorporates all the elements discussed in ParleE, organizing them in a more sequential and clearer manner Additionally, GRACE utilizes the MBTI personality model, which outlines more flexible personality traits.
The Greta model, proposed by Poggi and Pelachaud in 2003, is compared to GRACE and dynamic belief networks, which closely align with emotional processes The versatility of GRACE is evident as it effectively incorporates various elements of these dynamic networks While Greta's model focuses solely on the agent's beliefs and goals in relation to the current situation, we suggest integrating internal cognitive states through multiple disciplines that consider various aspects of an individual's life influencing emotional behaviors Additionally, we aim to introduce a new degree of flexibility in emotional responses, taking into account not only the situation but also the individual's personality traits.
Figure 13 - GRACE et les réseaux dynamique de Greta
Ensuite, on va voir le rapport en GRACE et EMA de Gratch et Marselle proposé en
In 2004, EMA provides a comprehensive overview of the decision-making process, aligning closely with the Behavior module of GRACE Additionally, EMA clearly identifies other elements that enhance our understanding of the functions within the GRACE modules However, it is important to note that personality factors have not yet been incorporated into the EMA model.
On termine cette partie de comparaison avec le modèle le plus récent GALAAD proposé par Carole Adam et ses collègues en 2005
This model focuses on dialogue scenarios, analyzing emotional aspects solely within the context of conversation, which can enhance the GRACE processes A unique feature of GRACE is its incorporation of personality, adding depth to behaviors and dialogues based on the individual's personal narrative within the game.
In order to address key aspects of the emotional process—including appraisal, coping, and personality—we conducted a comparison of recent computational models as detailed in the current state of the art.
Table 1 - Synthèse de comparaison des modèles calculatoires récents
Non Modèle appraisal coping personality
2 ParleE [6] Oui Pas abordé Rousseau’s model[7]
3 Robot Kismet [8] Oui Pas abordé Non
4 Greta [9] Oui Pas abordé Personality trait
A notable feature of GRACE is its unique Intuition module, which was not available in previous versions This module has the potential to enhance emotional behaviors, making them more flexible, understandable, and unpredictable.
4 Objectif et validation d’une instance de GRACE
Objectif de l’expérimentation
The primary objective is to validate the relevance of each module in the proposed model However, implementing the entire generic model requires significant time investment During this internship, a simplified demonstration could be conducted to simulate the model's reduced functionality The goal of this demonstration is to assess whether the model effectively captures changes in emotional states.
To achieve our objective, we have decided to conduct a demonstration simulating the functionality of two modules: Behavior and Body For the remainder, we will manually address the interpretation of events and the internal cognitive state related to desires In this initial demonstration, we will not cover physiological interpretation or emotions Regarding personality, we aim to implement the Energy dimension (Extraversion/Introversion) in our first demonstration, showcasing two instances that simulate different personality types: one extraverted and the other introverted.
Scénario de validation
The GRACE demo will be tested with individuals unfamiliar with its graphical display The experimental scenario is outlined as follows:
• 1) On prend quelques personnes pour jouer un rôle d’observateur de GRACE en leur faisant regarder l’affichage graphique
We will present a series of events to which GRACE will respond in the presence of observers In this instance, the events will be derived from the listening of a children's story that the observers will hear.
• 3) L’instance de GRACE réagit dynamiquement à ces événements Les observateurs regardent ce qui se passe sur l’affichage graphique
• 4) Après quoi, on demande aux observateurs de remplir un questionnaire pour savoir ce qu’ils voient sur l’affichage graphique et de quoi elle s’agit selon leur opinion
Our hypothesis is that if individuals can recognize what we present on the graphical display, which visually represents GRACE's emotional state in relation to each event, it indicates that the model is valid.
Fonctionnement de l’instance réduite de GRACE
Dans cette première instance du modèle, nous voulons implémenter les deux composants Behavior et Body, en fait c’est ce qui est en gras dans la figure 16 :
Figure 16 - Fonctionnement de l'instance réduite de GRACE
While the goal is to validate the Body component, which represents the phase of emotional expression, we also aim to implement the Behavior component This will allow us to explore how this component processes information received from cognitive interpretation.
Cognitive interpretation involves an internal cognitive process where data is manually adjusted based on personality types Extraversion enhances data value, while introversion reduces it Additionally, the functioning of the selected phases, Behavior and Body, is streamlined Simple formulas will be used to calculate Behavior, determining the intensity of emotional responses Initial trials aim to identify effective methods of expression.
Données en entrées
Nous avons dộcidộ d’utiliser l’histoire ô trois petits cochons ằ comme le suite de l’événement en entrée
In this analysis, we will explore the narrative to uncover the storyteller's intent This involves examining the story through key questions: What interpretations did the storyteller aim to convey at specific moments? What emotions were intended to resonate with the audience?
The audio file of the story features a recording by Sabine Letellier-Zarshenas, focusing on the segment about the second pig The text of this segment is transcribed from the tale "The Three Little Pigs," included in the second CD of "100 Tales, Fables, and Short Stories," published by Eveil et Découverte under exclusive license from EMI Music France.
- L’histoire est découpée en chunks 2 , chaque chunk est associé des valeurs d’interprétation (urgence, danger, affection) La valeur de Urgence et
In the GRACE model, the danger value ranges from 0 to 1, while the affective value varies between -1 and 1, reflecting a spectrum from worse to better This interpretative phase is crucial for the model's functionality, as it provides essential information for the consciousness process to effectively process data and generate appropriate emotional responses The selection of three parameters—urgency, danger, and affection—aims to measure incoming events to produce output values for three basic emotions: joy, sadness, and fear, along with excitement The calculation of excitement is based on the level of urgency of the event and the individual's emotional state.
2 Un chunk s’agit d’un groupe des mots qui est évalué comme un nouvel événement
In the internal cognitive state, the sentiment module is omitted, focusing instead on estimating future desires linked to various events By analyzing past occurrences, the internal cognitive state aims to predict potential future events and how emotions may evolve in response to these anticipated outcomes This predictive process highlights the relationship between past experiences and future emotional responses.
Mise en œuvre
Programmation
L’histoire d’entrée est découpé manuellement en chunks (chaque chunk correspond à un événement) Chaque chunk va être estimé un terme de trois paramètres que l’on a choisi : urgence, danger et affection
- Urgence Urg e : pour mesurer l’impératif de l’événement Cela est pour mesurer la rapidité de la réponse de la personne par rapport à l’événement détecté
- Danger Dage: pour mesurer le niveau de danger de l’événement au héro auquel la personne fait attention Le héro dans ce scénario est les cochons dans l’histoire
- Affection Aff e : ce paramètre a pour but de mesurer la favorabilité de l’événement à GRACE Il est positif si l’événement est en faveur ou négatif si l’événement est défavorable au personnage
Les données traitées sont enregistrées dans des fichiers de texte [Veuillez voir l’annexe pour un exemple de ces fichiers]
Before being processed through the Behavior module, interpretation data will be adjusted based on personality type As mentioned earlier, our demonstration features two instances representing two personality types: extraversion and introversion The extraversion instance is referred to as Kel.
Kel amplifie la valeur des données d’interprétation en appliquant la formule
Ly modifie la valeur des données d’interprétation en appliquant la formule suivante :
4.5.1.2 Données de l’état cognitif interne
In this initial demonstration, the Internal Cognitive State component solely predicts the changes in basic emotions of the simulated individual in response to each detected event This prediction is manually assessed, and the data is stored in a text file as input for the program For a concrete example of these predictions, please refer to the appendix.
The internal cognitive state is responsible for the temporal fluctuations of a person's spontaneous basic emotions These emotions can increase or decrease in response to detected events, with joy rising in reaction to positive occurrences and fear escalating in response to perceived threats Each basic emotion changes according to a normal distribution, characterized by distinct means and standard deviations for each emotion.
De notre avis, la peur varie beaucoup plus vite que la tristesse et la joie La tristesse varie par contre moins vite que la joie
Les intensités des émotions de base en sortie seront calculées dans le module
Initially, we aim to implement simplified formulas designed to calculate the intensity of basic emotions based on the description of current events and the prediction of internal cognitive states.
Pour chaque événement détecté, on calcule :
• Peurcourant = Peurprécédante + ( Affper * Peurprevue ) + ( 0.1 * Dagper * Urgper )
Figure 17 - Affichage graphique du démo
The three circles illustrate the current state of each basic emotion—joy (in green), sadness (in gray), and fear (in red)—at any given moment, reflecting their dynamic changes over time This fluctuation simulates the emotional state of an individual, particularly in relation to the story of "The Three Little Pigs," which is believed to significantly impact these core emotions While these three emotions are highlighted, there is potential to incorporate additional basic emotions in future iterations to provide a more comprehensive representation of the emotional state.
Expérimentation
L’expérimentation est en cours d’exécution…
Initial tests with participants have yielded promising results, but further testing is necessary to gather a larger pool of feedback for a comprehensive evaluation of the instance's functionality and the relevance of the display component (Body) Among the 15 individuals who took part in the experiment, 50% reported noticing what was displayed on the screen.
Cela nous donne un premier signal positif pour continuer notre développement du modèle GRACE.