In mid-2023, we performed an experiment in autonomous buses in Stockholm, Sweden, to evaluate the role that social robots might have in such settings, and their effects on passengers' feeling of safety and security, given the absence of human drivers or clerks. To address the situations that may occur in autonomous public transit (APT), we compared an embodied agent to a disembodied agent. In this video publication, we showcase some of the things that worked with the interactions we created, and some problematic issues that we had not anticipated.

This thesis addresses the topic of grounding in human-robot interaction, that is, the process by which the human and robot can ensure mutual understanding. To explore this topic, the scenario of a robot holding a presentation to a human audience is used, where the robot has to process multimodal feedback from the human in order to adapt the presentation to the human's level of understanding.

First, the use of behaviour trees to model real-time interactive processes of the presentation is addressed. A system based on the behaviour tree architecture is used in a semi-automated Wizard-of-oz experiment, showing that audience members prefer an adaptive system to a non-adaptive alternative.

Next, the thesis addresses the use of knowledge graphs to represent the content of the presentation given by the robot. By building a small, local knowledge graph containing properties (edges) that represent facts about the presentation, the system can iterate over that graph and consistently find ways to refer to entities by referring to previously grounded content. A system based on this architecture is implemented, and an evaluation using simulated users is presented. The results show that crowdworkers comparing different adaptation strategies are sensitive to the types of adaptation enabled by the knowledge graph approach.

In a face-to-face presentation setting, feedback from the audience can potentially be expressed through various modalities, including speech, head movements, gaze, facial gestures and body pose. The thesis explores how such feedback can be automatically classified. A corpus of human-robot interactions is annotated, and models are trained to classify human feedback as positive, negative or neutral. A relatively high accuracy is achieved by training simple classifiers with signals found mainly in the speech and head movements.

When knowledge graphs are used as the underlying representation of the system's presentation, some consistent way of generating text, that can be turned into speech, is required. This graph-to-text problem is explored by proposing several methods, both template-based and methods based on zero-shot generation using large language models (LLMs). A novel evaluation method using a combination of factual, counter-factual and fictional graphs is proposed. 

Finally, the thesis presents and evaluates a fully automated system using all of the components above. The results show that audience members prefer the adaptive system to a non-adaptive system, matching the results from the beginning of the thesis. However, we note that clear learning results are not found, which means that the entertainment aspects of the presentation are perhaps more prominent than the learning aspects.

Denna avhandling behandlar ämnet multimodal kommunikativ grundning (grounding) mellan robotar och människor. Detta är processen för hur en människa och en robot kan säkerställa att de har en gemensam förståelse. För att utforska detta ämne ämne, används ett scenario där en robot håller en presentation för en mänsklig publik. Roboten måste analysera multimodala signaler från människan för att anpassa presentationen till människans nivå av förståelse.

Först undersöks hur beteendeträd kan användas för att modellera realtidsaspekterna av interaktionen mellan robotpresentatören och dess publik. Ett system som baseras på beteendeträdsarkitekturen används i ett delvis automatiskt, delvis människostyrt experiment, där det visas att publikmedlemmar i labbmiljö föredrar ett system som anpassar presentationen till deras reaktioner över ett som inte anpassar sin presentation.

Efter detta, urdersöker också avhandlingen hur kunskapsgrafer kan användas för att representera innehållet som roboten presenterar. Om en liten, lokal kunskapsgraf byggs så att den innehåller relationer (kanter) som representerar fakta i presentationen, så kan roboten iterera över grafen och konsekvent hitta refererande uttryck som använder sig av kunskap som publiken redan har. Ett system som baseras på denna arkitektur implementeras, och ett experiment med simulerade interaktioner utförs och presenteras. Experimentets resultat visar att utvärderare som jämför olika anpassningsstrategier föredrar ett system som kan utföra den sortens anpassning som grafmetoden tillåter. 

Publikens reaktioner i ett presentationsscenario kan ske genom olika modaliteter, som tal, huvudrörelser, blickriktning, ansiktsuttryck och kroppsspråk. För att klassificera kommunikativ återmatning (feedback) av dessa modaliteter från presentationspubliken, utforskas hur sådana signaler kan analyseras automatiskt. En datamängd med interaktioner mellan en människa och vår robot annoteras, och statistiska modeller tränas för att klassificera mänskliga återmatningssignaler från flera olika modaliteter som positiva, negativa eller neutrala. En jämförelsevis hög klassifikationsprecision uppnås genom att träna enklare klassifikationsmodeller på relativt få klasser av signaler i tal- och huvudrörelsemodaliteterna. Detta antyder att museiscenariot med en robotpresentatör inte uppmuntrar publiken att använda komplicerade, mångtydiga kommunikativa beteenden.

När kunskapsgrafer används som presentationssystemets informationsrepresentation, behövs det konsekventa metoder för att generera text som kan omvandlas till tal, från grafdata. Graf-till-text-problemet utforskas genom att föreslå flera olika metoder, både enklare mall-baserade sådana och mer avancerade metoder baserade på stora språkmodeller (LLM:er). Genom att föreslå en ny utvärderingsmetod där sanna, fiktiva och falska grafer genereras, visar vi också att sanningshalten i vad som uttrycks påverkar kvaliteten i texten som LLM-metoderna ger från kunskapsgrafdata.

Avhandlingen använder sig slutligen av alla de ovanstående föreslagna komponenterna i ett och samma helautomatiska presentationssystem. Resultaten visar att publikmedlemmar föredrar ett system som anpassar sin presentation över ett som inte anpassar sin presentation, vilket speglar resultaten från början av avhandlingen. Vi ser också att tydliga inlärningsresultat uteblir i detta experiment, vilket kanske kan tolkas som att publikmedlemmarna i museiscenariot snarare letar efter en underhållare än efter en lärare som presentatör.

An interesting application for social robots is to act as a presenter, for example as a museum guide. In this paper, we present a fully automated system architecture for building adaptive presentations for embodied agents. The presentation is generated from a knowledge graph, which is also used to track the grounding state of information, based on multimodal feedback from the user. We introduce a novel way to use large-scale language models (GPT-3 in our case) to lexicalise arbitrary knowledge graph triples, greatly simplifying the design of this aspect of the system. We also present an evaluation where 43 participants interacted with the system. The results show that users prefer the adaptive system and consider it more human-like and flexible than a static version of the same system, but only partial results are seen in their learning of the facts presented by the robot.

People often form immediate expectations about other people, or groups of people, based on visual appearance and characteristics of their voice and speech. These stereotypes, often inaccurate or overgeneralized, may translate to robots that carry human-like qualities. This study aims to explore if nationality-based preconceptions regarding appearance and accents can be found in people's perception of a virtual and a physical social robot. In an online survey with 80 subjects evaluating different first-language-influenced accents of English and nationality-influenced human-like faces for a virtual robot, we find that accents, in particular, lead to preconceptions on perceived competence and likeability that correspond to previous findings in social science research. In a physical interaction study with 74 participants, we then studied if the perception of competence and likeability is similar after interacting with a robot portraying one of four different nationality representations from the online survey. We find that preconceptions on national stereotypes that appeared in the online survey vanish or are overshadowed by factors related to general interaction quality. We do, however, find some effects of the robot's stereotypical alignment with the subject group, with Swedish subjects (the majority group in this study) rating the Swedish-accented robot as less competent than the international group, but, on the other hand, recalling more facts from the Swedish robot's presentation than the international group does. In an extension in which the physical robot was replaced by a virtual robot interacting in the same scenario online, we further found the same results that preconceptions are of less importance after actual interactions, hence demonstrating that the differences in the ratings of the robot between the online survey and the interaction is not due to the interaction medium. We hence conclude that attitudes towards stereotypical national representations in HRI have a weak effect, at least for the user group included in this study (primarily educated young students in an international setting).

In any system that uses structured knowledgegraph (KG) data as its underlying knowledge representation, KG-to-text generation is a useful tool for turning parts of the graph data into text that can be understood by humans. Recent work has shown that models that make use of pretraining on large amounts of text data can perform well on the KG-to-text task, even with relatively little training data on the specific graph-to-text task. In this paper, we build on this concept by using large language models to perform zero-shot generation based on nothing but the model’s understanding of the triple structure from what it can read. We show that ChatGPT achieves near state-of-the-art performance on some measures of the WebNLG 2020 challenge, but falls behind on others. Additionally, we compare factual, counter-factual and fictional statements, and show that there is a significant connection between what the LLM already knows about the data it is parsing and the quality of the output text.

The work presented here is a culmination of developments within the Swedish project COIN: Co-adaptive human-robot interactive systems, funded by the Swedish Foundation for Strategic Research (SSF), which addresses a unified framework for co-adaptive methodologies in human-robot co-existence. We investigate co-adaptation in the context of safe planning/control, trust, and multi-modal human-robot interactions, and present novel methods that allow humans and robots to adapt to one another and discuss directions for future work.

Intelligent agents interacting with humans through conversation (such as a robot, embodied conversational agent, or chatbot) need to receive feedback from the human to make sure that its communicative acts have the intended consequences. At the same time, the human interacting with the agent will also seek feedback, in order to ensure that her communicative acts have the intended consequences. In this review article, we give an overview of past and current research on how intelligent agents should be able to both give meaningful feedback toward humans, as well as understanding feedback given by the users. The review covers feedback across different modalities (e.g., speech, head gestures, gaze, and facial expression), different forms of feedback (e.g., backchannels, clarification requests), and models for allowing the agent to assess the user's level of understanding and adapt its behavior accordingly. Finally, we analyse some shortcomings of current approaches to modeling feedback, and identify important directions for future research.

Feedback is an essential part of all communication, and agents communicating with humans must be able to both give and receive feedback in order to ensure mutual understanding. In this paper, we analyse multimodal feedback given by humans towards a robot that is presenting a piece of art in a shared environment, similar to a museum setting. The data analysed contains both video and audio recordings of 28 participants, and the data has been richly annotated both in terms of multimodal cues (speech, gaze, head gestures, facial expressions, and body pose), as well as the polarity of any feedback (negative, positive, or neutral). We train statistical and machine learning models on the dataset, and find that random forest models and multinomial regression models perform well on predicting the polarity of the participants' reactions. An analysis of the different modalities shows that most information is found in the participants' speech and head gestures, while much less information is found in their facial expressions, body pose and gaze. An analysis of the timing of the feedback shows that most feedback is given when the robot makes pauses (and thereby invites feedback), but that the more exact timing of the feedback does not affect its meaning.

In this paper, we address the problem of how an interactive agent (such as a robot) can present information to an audience and adaptthe presentation according to the feedback it receives. We extend a previous behaviour tree-based model to generate the presentation from a knowledge graph (Wikidata), which allows the agent to handle feedback incrementally, and adapt accordingly. Our main contribution is using this knowledge graph not just for generating the system’s dialogue, but also as the structure through which short-term user modelling happens. In an experiment using simulated users and third-party observers, we show that referring expressions generated by the system are rated more highly when they adapt to the type of feedback given by the user, and when they are based on previously grounded information as opposed to new information.

In dialogue, speakers continuously adapt their speech to accommodate the listener, based on the feedback they receive. In this paper, we explore the modelling of such behaviours in the context of a robot presenting a painting. A Behaviour Tree is used to organise the behaviour on different levels, and allow the robot to adapt its behaviour in real-time; the tree organises engagement, joint attention, turn-taking, feedback and incremental speech processing. An initial implementation of the model is presented, and the system is evaluated in a user study, where the adaptive robot presenter is compared to a non-adaptive version. The adaptive version is found to be more engaging by the users, although no effects are found on the retention of the presented material.