Embodiment and Computational Creativity (PDF)

Embodiment and Computational Creativity Christian Guckelsberger1–3,†, Anna Kantosalo2,†, Santiago Negrete-Yankelevich4,‡ and Tapio Takala2,‡ 1Finnish Center for Artificial Intelligence 2Department of Computer Science, Aalto University, Espoo, Finland 3School of Electronic Engineering and Computer Science, Queen Mary University of London, London, UK 4 Department of Information Technology, Metropolitan Autonomous University (Cuajimalpa), Mexico City, Mexico †Contributed equally and share first authorship ‡Contributed equally and share second authorship [email protected], [email protected], [email protected], [email protected] Abstract We conjecture that creativity and the perception of cre- ativity are, at least to some extent, shaped by embod- iment. This makes embodiment highly relevant for computational creativity (CC) research, but existing re- search is scarce and the use of the concept highly am- biguous. We overcome this situation by means of a systematic review and a prescriptive analysis of pub- lications at the International Conference on Computa- tional Creativity. We adopt and extend an established typology of embodiment to resolve ambiguity through identifying and comparing different usages of the con- cept. We collect, contextualise and highlight opportuni- ties and challenges in embracing embodiment in CC as a reference for research, and put forward important direc- tions to further the embodied CC research programme. Introduction Most researchers agree that creativity and intelligence are closely intertwined (Kaufman and Plucker, 2011). More- over, it is widely accepted that intelligence is conditioned on embodiment (Brooks, 1991; Clark, 1998); with the words of Pfeifer and Scheier, “intelligence cannot merely exist in the form of an abstract algorithm but requires a physical instan- tiation, a body” (Pfeifer and Scheier, 2001, p. 649). Without insisting on physicality, we conjecture that creativity, as a form of intelligent cognition, is also shaped by embodiment. This makes embodiment highly relevant across the en- tire continuum of computational creativity (CC) research (P´erez y P´erez, 2018; Veale, Cardoso, and P´erez y P´erez, 2019), from engineering artificial systems that can be con- sidered autonomously creative (Colton, 2008; Colton and Wiggins, 2012), to understanding creativity in living be- ings through computational modelling and simulation (Bo- den, 2003). Here, some of the most striking questions are if reproducing human-like creativity by computational means is at all possible without also reproducing human embod- iment (Guckelsberger, Salge, and Colton, 2017; Valverde- P´erez and Negrete-Yankelevich, 2018), and how changes to a system’s embodiment affect its potential creativity. Unlike psychologists, whose human subjects share simi- lar embodiments, CC researchers can go beyond investigat- ing the effect of embodiment on creativity per se and ex- plore how people’s perception of creativity (Colton, 2008; Colton, Pease, and Saunders, 2018) as exhibited by an artifi- cial system is affected by their own and the system’s embod- iment. If such an effect existed, embodiment would have to be considered an integral factor in the evaluation of CC (Jor- danous, 2012) to facilitate fairer comparisons between com- putational systems and with human creativity. Moreover, using this knowledge, researchers could tune a system’s em- bodiment to improve the perception of its creativity, and, vice versa, how it perceives the creativity of others. Given these potential ramifications, it is surprising and alarming that “embodiment” seems to have received little at- tention in CC research. One potential reason for the apparent void is that the very concept is highly ambiguous (Ziemke, 2003). Theories of embodied cognition – from minimal ac- counts that “rule out anatomy and bodily movement as im- portant” (Gallagher, 2011), to radical approaches that under- stand cognition as inextricably bound to bodily processes – conceptualise embodiment differently. Moreover, embod- ied cognition is closely associated with other, popular extra- cranial and extra-bodily theories of cognition, in particular theories of enactive, embedded and extended cognition. Due to their proximity and interdependence, they are frequently grouped together into the complex of “4E cognition”. Crucially though, there appears to be very little awareness of this ambiguity and complexity within CC research. At present, it is perfectly imaginable that two researchers excit- edly referred to the “embodiment” of their respective system without noticing that they are talking about entirely differ- ent things. We consider this a major problem, given that the various types of embodiment likely have radically different effects on (the perception of) creativity. We argue that the advancement of the field through concerted investigations requires the use of common definitions of embodiment. The first goal of this paper is to counteract this ambigu- ity and provide a rich overview of what some authors have already coined “embodied computational creativity (CC)” (Saunders and Gemeinboeck, 2014; Guckelsberger, Salge, and Colton, 2017; Colton, Pease, and Saunders, 2018) re- search. Based on a systematic review of related work at the International Conference on Computational Creativity (ICCC) as the prime and domain-agnostic venue of CC re- search, we answer the following research questions: RQ1: What types of embodiment have been embraced in CC research, and how has the usage evolved over time? Proceedings of the 12th International Conference on Computational Creativity (ICCC ’21) ISBN: 978-989-54160-3-5 192  RQ2: Why did CC researchers embrace these embodiment types in their work, and what challenges did they face? RQ3: What does CC research reveal about the relationship of embodiment and (the perception of) creativity? To counteract ambiguity in the usage of the embodiment concept, we extend and apply a well established typology of embodiment informed by cognitive science to assess the specific types addressed in each relevant contribution. By making transparent which types of embodiment have been embraced, and by highlighting our challenges in assess- ing them, we want to provide a frame of reference for re- searchers to adequately and unambiguously address ques- tions of embodiment in their work. Our insights moreover allow us to to provide recommendations for an embodied CC research programme – the second goal of this paper. Types of Embodiment Much research focuses on distinguishing theories of embod- ied cognition (e.g. Gallagher, 2011), but comparisons of the underlying and varying conceptualisations of embodiment are rare. To disambiguate different uses of the embodiment concept in CC, we adopt and extend the well-established ty- pology by Ziemke (2003), who distinguishes six types of embodiment informed by research in cognitive science and robotics. We introduce three additions to this typology (one additional type, two additions to existing types) based on more recent insights, and highlight them in italics below. structural coupling, characterising systems that can per- turbate, and, vice versa, be perturbated by their surround- ing environment (Varela, Rosch, and Thompson, 1991). Such perturbations facilitate a minimal interaction be- tween the system and environment, in which each has the potential to affect the other’s state (Quick et al., 1999). historical, characterising systems whose present state is the result of a history of structural couplings, devel- oped through interactions with the environment over time (Varela, Rosch, and Thompson, 1991; Ziemke, 1999). virtual, characterising simulated systems embedded in and distinguished from a simulated environment. The virtual body can act on the environment and vice versa. physical, characterising systems with a physical body (Brooks, 1990; Pfeifer and Scheier, 2001) that can inter- act with the environment by being subjected to and by exercising physical force. Most prominently, robots are physically embodied (Pfeifer, Iida, and Bongard, 2005). organismoid, characterising virtually or physically em- bodied systems with the same or a similar shape and sen- sorimotor equipment as living organisms, e.g. animals. We consider humanoid embodiment as approximations of the human body a subset of organismoid embodiment. organismic, applying to living and artificial systems capa- ble of organisational closure, i.e. of maintaining their or- ganisation and surrounding boundary against internal and external perturbations by means of self-producing pro- cesses (Von Uexk¨ull, 1920; Maturana and Varela, 1987). A prominent, minimal example is the living cell which, in a self-referential process, maintains its organisation, in- cluding its membrane, against perturbations from the sur- rounding environment (Agmon, Gates, and Beer, 2016). We briefly justify our additions. We have complemented physical with virtual embodiment, because AI researchers have successfully reproduced (super-)human cognitive abil- ities in virtual agents, embedded in e.g. high fidelity physics simulations (Lillicrap et al., 2015) or, often more coarse, videogame worlds (Mnih et al., 2015). Applying AI tech- niques to virtual agents in simulated worlds rather than to physically embodied systems allows for scalability, incre- mental development, and rapid iteration, amongst other ad- vantages (Kiela et al., 2016). As a corollary, we have ex- tended organismoid embodiment to virtually embodied sys- tems, e.g. in the form of believable game characters with a human or animal-like appearance. We have finally extended organismic embodiment to artificial systems. Originally re- stricted to the biochemical domain, this type required the capacity for autopoiesis, i.e. self-production, to facilitate a radical form of autonomy. Varela overcomes this limitation by introducing the concept of organisational closure as “op- erational characterization of autonomy in general, living or otherwise” (Varela, 1979). Froese and Ziemke (2009) advo- cate that organisational closure can be realised by AI sys- tems, and survey existing examples. We can thus consider organismic embodiment in artificial systems, which makes it relevant for CC research. Although organismic embodi- ment is arguably the least well-established type, we include it for its presence in existing CC theory (Saunders, 2012; Guckelsberger, Salge, and Colton, 2017), its potential future implications for CC, and its central role in related debates, e.g. on agency (Polani, Ikegami, and Biehl, 2016). In our adaptation of Ziemke’s (2003) typology, we have dropped what e.g. Dautenhahn (1997) and Barsalou et al. (2003) refer to as social embodiment, because it denotes the use of different types of embodiment to facilitate social in- teraction (Ziemke, 2003), and is thus orthogonal to, and not at the same “atomic” level, as the other types. Metzinger (2014) has proposed a distinction between 1st, 2nd and 3rd order embodiment based on a system’s computational abili- ties, corresponding to (1) physical, reactive systems without explicit computation, (2) systems that explicitly represent themselves as embodied agents, and (3) systems that can consciously experience some of these body representations. We disregard this typology as (i) it is not derived from exist- ing work on embodied cognition more generally and serve the specific purpose of grounding (artificial) consciousness as one aspect of cognition, and because (ii) it would only warrant little differentiation of existing work; most CC sys- tems presently fall into a gap between the 1st and 2nd type. Crucially, the extended typology is only loosely hierar- chical. Historically as well as physically and virtually em- bodied systems are all structurally coupled. Physically and virtually embodied systems in turn can, but do not have to be historically embodied. Organismoid systems can be vir- tually or physically embodied, but we reserve organismic embodiment to physically embodied systems, as autonomy via organisational closure relies on physical forces that pose a real threat to a system’s organisation and boundary. Proceedings of the 12th International Conference on Computational Creativity (ICCC ’21) ISBN: 978-989-54160-3-5 193  Review Method To investigate how embodiment has been discussed in CC research, we performed a systematic literature review on the proceedings of the International Conference on Computa- tional Creativity (ICCC), the prime venue for CC research, between its inception in 2010 and its latest edition in 2020. We expect our findings to be representative, as ICCC gath- ers a wide audience of CC researchers and practitioners, and welcomes contributions covering any creative domain, cre- ative practice and aspect of creative cognition (Association for Computational Creativity, 2021). We constrained our re- view to paper candidates that explicitly mention the words “embodiment”, “embodied”, “disembodiment”, “disembod- ied”, “embody”, or “embodying”. We acknowledge that our reliance on the explicit usage of the word may overlook a large amount of potentially rel- evant papers. This particularly concerns work on robotics, which often does not include explicit mentions of the em- bodiment of the investigated systems. Simply including the term “robotics” in our search however was impractical, as, for fairness, it would have required to also include any other type of system characterised by the remaining embodiment types. Given the inclusiveness of some types, the amount of potentially related papers would likely go beyond what could reasonably be reviewed in depth. Moreover, we be- lieve that our present approach allows us to identify inten- tional and thus more informative discussions of the relation- ship between embodiment and (the perception of) creativity. For our initial candidate paper selection, we divided the past ICCC proceedings into pdf files, each containing one paper, short paper, demo description, or other peer reviewed publication. We used a wildcard search in Adobe Acro- bat reader with the phrase “*embod*” and the hyphenated version “*em-bod*”. In total we found 99 papers men- tioning the word “embodiment”, “embodied”, “disembodi- ment”, “disembodied”, “embody”, or “embodying” explic- itly, with a total of 491 matches to the search phrases. We then reduced the candidate papers for our final, qual- itative analysis in a two-stage process. We first excluded papers which mentioned these words only in the References section (7 papers), or as part of a general list of CC related topics (2 papers). We secondly excluded papers that used these keywords in a merely metaphorical way, such as sug- gesting that a specific algorithm or system “embodies” cer- tain values (51 papers). This left us with a final pool of 40 papers for in-depth analysis. Figure 1 illustrates the overall usage of the term over the years, thus partly answering RQ1. We assessed the embodiment of any system described in the remaining papers, either introduced there or through ref- erence to other work, based on Ziemke’s (2003) extended typology. Since explicit definitions of embodiment were mostly absent, this usually required us to look at the spe- cific characteristics of the system in question. To answer our research questions, we also gathered notes on: • Which challenges did we encounter in assessing the em- bodiment described in the paper? • Did we identify any embodiment types that were not yet present in our typology? (RQ1) 0 10 20 30 40 50 60 70 80 90 ICCC 2010 ICCC 2011 ICCC 2012 ICCC 2013 ICCC 2014 ICCC 2015 ICCC 2016 ICCC 2017 ICCC 2018 ICCC 2019 ICCC 2020 Papers with no mention Metaphorical use of embod* Non-metaphorical use of embod* Figure 1: Absolute annual numbers of ICCC papers using words derived from ‘embodiment’ in a (non-)metaphorical way. On average, 7.1% of papers use the word ‘embodi- ment’ or its derivatives non-metaphorically, with the lowest proportion (0%) in 2012 and the highest (11.8%) in 2017 and 2020. Overall the term is used throughout the proceedings, but non-metaphorical usage has increased over the years. • Why have the authors embraced this particular type of embodiment, and what challenges did they face? (RQ2) • What did the paper express about the relationship between embodiment and creativity, or between embodiment and the perception of creativity? (RQ3) Each of the remaining 40 papers was first assessed by one of the four researchers participating in the effort. We then discussed the challenges encountered, and cross-checked each paper with another colleague in the team. Any poten- tial disagreements were resolved in dedicated discussions. Another four papers were removed during this step because the explicit mentions to embodiment were cursory. This re- duced the final number of papers to 36. The answers to the questions presented above were anal- ysed by building an affinity diagram, a qualitative data anal- ysis method used to group data into emergent categories (Rogers, Sharp, and Preece, 2011, p. 286). We first built individual affinity diagrams from the answers collected for each question by grouping similar items together and la- belling them. We then merged categories between the differ- ent questions, yielding one large affinity diagram as a con- nected view of different aspects of embodiment in CC. We finally identified overarching themes within this structure. Findings We report our findings on the state-of-the-art of embodied CC in individual subsections. We first detail the types of embodiment identified in existing work, thus conclusively answering RQ1. We elaborate on our difficulties in this pro- cess later in the discussion section, where we go beyond the present findings and make recommendations for future re- search. Our review uncovered that existing insights on the relationship of embodiment and (the perception of) creativ- ity are often tightly aligned with researchers’ motivations to embrace a certain type of embodiment in their work. In the second part of our findings, we consequently address RQ2 and RQ3 jointly through themes corresponding to opportu- nities and challenges for embodied CC. We distinguish each Proceedings of the 12th International Conference on Computational Creativity (ICCC ’21) ISBN: 978-989-54160-3-5 194  Embodiment Type Authors Year Abbreviated Title Concrete System(s) Mult. Embodiments Struct. Coupling Historical Virtual Physical Organismoid Humanoid Organismic de Melo & Gratch 2010 Evolving Expression of Emotions — — � � � � � � � Saunders et al. 2010 Curious Whispers � — � � � � �� � � Kirsh 2011 Creative Cognition in Choreography — — � � � � � � � Gemeinboeck & Saunders 2013 Creative Machine Performance: Robotic Art � — � � � � �� � � Saunders, Chee & Gemeinboeck 2013 Evaluating Human-Robot Interaction � — � � � � �� � � Schubert & Mombaur 2013 Motion Dynamics in Abstract Painting � — � � � � � � � Schorlemmer et al. 2014 COINVENT: Concept Invention Theory — — � � � � � � � Davis et al. 2014 Building Artistic Computer Colleagues � — � � � � � � � Jacob & Magerko 2015 Interaction-based Authoring � — � � � � � � � Takala 2015 Preconceptual Creativity � — � � � � � � � Carlson et al. 2016 Cochoreo: Generative Choreography � — � � � � � � � McCormack & d’Inverno 2016 Designing Improvisational Interfaces � — � � � � � � � Crnkovic-Friis & Crnkovic-Friis 2016 Generative Choreography using Deep Learning � — � � � � � � � Guckelsberger et al. 2016 Supportive and Antagonistic Behaviour � � � � � � � � � Brown 2016 Understanding Musical Practices � — � � � � � � � Augello et al. 2017 Creative Robot Dance With Variational Encoder � — � � � � � � � Fitzgerald, Goel & Thomaz 2017 Human-Robot Co-Creativity: Task Transfer � — � � � � � � � Guckelsberger, Salge & Colton 2017 Non-Anthropocentric Intentional Creative Agency � � � � � � � � � Singh et al. 2017 Unified Classification and Generation Networks � — � � � � � � � Colton, Pease & Saunders 2018 Issues of Authenticity in Creative Systems � � � � � � � � � Hedblom et al. 2018 Under the Super-Suit: Conceptual Blending � — � � � � � � � Saunders & Gemeinboeck 2018 Performative Body Mapping for Robots � � �� � � � � � � Wicke & Veale 2018 Interview With the Robot � — � � � � � � � Jacob et al. 2019 Affordance-Based Generation of Object Variants � — � � � � � � � Onate, Mendez & Gervas 2019 Elevator Conversations Based on Emotions � — � � � � � � � Veale, Wicke & Mildner 2019 Duets Ex Machina: “Double Acts” � � � � � � � � � Loesel, Mirowski & Mathewson 2020 Do Digital Agents Do Dada? � � �� � � � � �� � Alexandre 2020 Creativity Explained by Cognitive Neuroscience � — � � � � � � � Chen et al. 2020 Breaking the Imitation Barrier � — � � � � � � � Ferguson et al. 2020 Automatic Similarity Detection in LEGO Ducks � — � � � � � � � Kantosalo et al. 2020 Modalities, Styles and Strategies — — � � � � � � � Kantosalo & Takala 2020 Five C’s for Human–Computer Co-Creativity � � � � � � � � � LaViers & Vidrin 2020 Can a Robot Do a Trust Fall? — — � � � � � � � Savery, Zahray & Weinberg 2020 Shimon the Rapper: Robot Rap Battles � — � � � � � � � Wallace et al. 2020 Movement Generation with Audio Features � — � � � � � � � Wicke & Veale 2020 Show, Don’t (Just) Tell � — � � � � � � � Table 1: Chronological overview of ICCC (2010-2020) papers mentioning the concept of embodiment explicitly and non- metaphorically. The circles � �, and � represent increasing degrees to which the described embodiment(s) match our types. Organismoid embodiment entails humanoid embodiment, but we also discriminate humanoid embodiment separately. For papers describing multiple embodied systems, e.g. a robot and a human, individual embodiment types were combined with a logic “or”. Papers introducing a “concrete system” have been marked; other papers consider theory or abstract systems. theme as a paragraph heading, point out sub-themes in bold, and highlight which embodiment type it relies on in italics. Embodiment Types Table 1 provides an overview of our final paper selection, together with our assessment of the described embodiment. We found that structural coupling and physical embodiment are most common, each appearing in almost twenty papers. Historical, virtual and organismoid mentions appear quite equally, each in about ten papers. Organismic embodiment was only identified in four papers; Guckelsberger, Salge, and Colton (2017) discuss it with respect to machines, while the other three instances relate to humans, which are organis- mically embodied by definition. Over the years, although there is considerable variation, the annual number of papers grows from 1.6 on average in the first half to 3.6 in the lat- ter half of the decade, until there is a sudden peak of 10 in 2020. Mentions of different types follow a similar trend, except that structural coupling gives way to the more spe- cific, physical embodiment in recent years. Out of the 36 papers, 23 present a concrete, embodied computational sys- tem, seven are at least partly theoretical or appealing to the embodiment of humans. In another seven papers, we were unable to identify one or multiple types of embodiment at Proceedings of the 12th International Conference on Computational Creativity (ICCC ’21) ISBN: 978-989-54160-3-5 195  all, indicated by a row of empty circles in Table 1. Our analysis highlights virtual and physical embodiment as strongest differentiators of existing work. It moreover identified a chasm between researchers poised to leverage humanoid embodiment, and those rejecting it for the benefits of other variations of organismoid embodiment. Through our review, we identified one embodiment type that was not directly present in our extension of Ziemke’s (2003) typology: Loesel, Mirowski, and Mathewson (2020) introduce “cyborg embodiment”, which bridges between virtual, organismoid (anthropomorphic) and physical em- bodiment. It was demonstrated in their theatrical experiment “AI Improv”, where a chatbot provides sentences to be artic- ulated by a human actor, and receives new prompts to react to from a backstage operator who monitors the on-stage di- alogue. The two people thus provide a split actuation and sensing interface to the artificial system, allowing for it to be embedded in the physical environment of the stage. Opportunities of Embodied CC Our analysis identified nine opportunities of embracing em- bodiment in CC. We can distinguish two sub-groups, based on how the themes relate to the concept of interaction. All types of embodiment distinguished in our typology assume at least a minimal form of interaction between an agent and its environment in the form of structural coupling. The first four themes concern how this embodiment-induced split provides opportunities for modelling a specific creative do- main, outsourcing computation, letting creativity emerge and stimulating it. The remaining five themes operate on a stronger notion of interaction between agents, i.e. where the embodied agent’s environment comprises interaction part- ners. In this group, embodiment is considered a means to model co-creativity, ground meaning, facilitate more natural interaction with people, support identification and empathy with the computational agent, and increase the CC system’s creative intentionality and autonomy. Domain Necessity Any type of embodiment presents the opportunity to model creative processes that unfold be- tween an agent and their environment. Some creative domains may necessitate this split more than others in or- der to comprehensively model the creative processes within. Dance choreography represents a prime example (Augello et al., 2017; Carlson et al., 2016), but embodiment has also been embraced in e.g. music (Schorlemmer et al., 2014) and painting (Schubert and Mombaur, 2013; Singh et al., 2017) to model creative processes that rely on sensorimotor feed- back between an agent and their surroundings. Outsourcing Computation Embodied CC has adopted several premises of embodied AI more generally, notably the use of physical embodiment to outsource computation into the physical world. Saunders, Chee, and Gemein- boeck (2013) note that physical embodiment allows artifi- cial agents “to take advantage of properties of the physical environment that would be difficult or impossible to simu- late computationally” (paraphrasing Brooks, 1990; see also Gemeinboeck and Saunders, 2013). They thus relate to one of the most prominently articulated benefits of embodied AI: the use of the world “as its own model” (Brooks, 1991). Emergent Creativity While the outsourcing of computa- tion seems a mere engineering benefit at first, it has ma- jor implications for the creativity that a physically embod- ied system can potentially exhibit. In particular, it allows to realise the very premise of systems theories of creativ- ity (Csikszentmihalyi, 1988), the emergence of creative behaviour through an agent’s interaction with their en- vironment, including other agents: “embodiment provides opportunities for agents to experience the emergence of ef- fects beyond the computational limits that they must work within” (Saunders, Chee, and Gemeinboeck, 2013). In the art installation Zwischenr¨aume, the robots’ creative agency “is not predetermined but evolves based on what happens in the environment they examine and manipulate” (Gemein- boeck and Saunders, 2013). This emergence benefits from a controller that is not pre-coded but sensitive to an agent’s changing embodiment. Several authors (e.g. Saunders et al., 2010; Saunders, Chee, and Gemeinboeck, 2013; Guckels- berger et al., 2016; Guckelsberger, Salge, and Colton, 2017) highlight the use of computational intrinsic motivation to this end. A system with a suitable controller can leverage its embodiment to expand its behavioural range beyond what can be anticipated by the system designer, realising novelty as core criterion for creativity (Rhodes, 1961). Stimulation of Creativity As a specific case of emer- gent creative behaviour, several researchers argue that con- straints imposed through embodiment can stimulate cre- ativity. Drawing on Pickering (2005), Saunders et al. high- light that the “world offers opportunities, as well as pre- senting constraints: human creativity has evolved to ex- ploit the former and overcome the latter, and in doing both, the structure of creative processes emerge” (Saunders et al., 2010). Guckelsberger et al. (2016) argue in an artistic context, drawing on different embodied CC systems and a thought-experiment on the physically embodied robot soci- ety Curious Whispers (Saunders et al., 2010), that overcom- ing embodiment-related constraints in an environment can necessitate and – given a suitable agent controller – yield creativity. Takala (2015) demonstrates this on a simulated robotic arm capable of inventing new and useful movements when encountering obstacles. This also highlights that cre- ative action is possible without creative reasoning, a dis- tinction which is later picked up by Fitzgerald, Goel, and Thomaz (2017). By utilising virtual embodiment, Takala demonstrates that the effect of embodiment constraints on creativity can be investigated without physical embodiment. However, the use of physical embodiment can better allevi- ate doubts about the emergent behaviour being truly novel, and not engineered a priori into a simulated environment. Co-Creativity Many of the analysed papers express a fo- cus on stronger forms of interaction between agents. Two particular modes of interaction are given by human-machine and machine-machine co-creativity (Saunders and Bown, 2015; Kantosalo and Toivonen, 2016). This focus can be ex- plained with the observation that embodiment is a prereq- Proceedings of the 12th International Conference on Computational Creativity (ICCC ’21) ISBN: 978-989-54160-3-5 196  uisite for co-creativity. Guckelsberger et al. (2016) high- light that “co-creative and social creativity systems are only meaningful if each agent has a different perspective on a shared world, allowing them to complement each other, and for creativity to emerge from their interaction”. The neces- sary separation of agent and environment is crucially facili- tated by any type of embodiment. This allows for the attribu- tion of embodiment to a given system, based on the systemic nature of the system alone. The next three themes represent additional lenses on human-machine co-creativity. Grounding Meaning Another central premise of more radical theories of embodied cognition, bordering to enac- tivism (Varela, Thompson, and Rosch, 2017), is that phys- ical embodiment can overcome symbolic representations and ground meaning in sensorimotor interaction (e.g. Dreyfus, 1992). Colton, Pease, and Saunders (2018) stress that this allows re-representing creative domains in action. As an example, they refer to the Marimba playing robot Shi- mon (Hoffman and Weinberg, 2010) which represents mu- sic as choreography of physical gestures. They also em- phasise the grounding of machine “life experiences” as an important factor in increasing the perception of authen- ticity in CC systems. Related, Wicke, Veale and Mildner exploit robot gestures to provide the illusion of grounding computer-generated stories (Wicke and Veale, 2018, 2020; Veale, Wicke, and Mildner, 2019), thus leveraging embodi- ment to affect the perception of their robot’s creativity. Natural Interaction Saunders et al. (2010) are first to stress that physical embodiment allows for CC systems to be embedded in rich social and cultural environments. This enables “computational agents to be creative in envi- ronments that humans can intuitively understand” (Saun- ders, Chee, and Gemeinboeck, 2013). Robotic art installa- tions are highlighted as one means to “gain access to shared social spaces with other creative agents, e.g., audience mem- bers” (Gemeinboeck and Saunders, 2013). Existing research often stresses that situating CC systems in physical space re- alises more natural interaction by established means, and can unleash new modes of interaction. This is explained by physical embodiment affording tight feedback loops (Wicke and Veale, 2018) and providing stronger cues to the human interaction partners (Saunders, Chee, and Gemein- boeck, 2013). In some instances, this interaction is con- strained to a few invididuals, e.g. when situating a robot on stage to interact with musicians (Savery, Zahray, and Wein- berg, 2020) or actors (Loesel, Mirowski, and Mathewson, 2020). In their art installation Zwischenr¨aume, Gemein- boeck and Saunders in contrast open the interaction to a wider audience in an exhibition space, permitting “the devel- opment of significantly new modes of interaction” and “en- gaging a broad audience in the questions raised by models of artificial creative systems” (Gemeinboeck and Saunders, 2013). Within natural interaction spaces, physical embod- iment can “improve the relationship between humans and AI, inspiring humans in new creative ways” (Savery, Zahray, and Weinberg, 2020), e.g. in partnering a human musician with Shimon’s reincarnation as rapper. Identification & Empathy Within the overarching theme of agent interaction, researchers embraced physical, and in particular organismoid and humanoid embodiment to fa- cilitate and improve communication and, consequently, to afford identification, empathy and affect between hu- man and robot, or within a society of robots. Gemeinboeck and Saunders highlight embodied action in their installation Zwischenr¨aume as a means for communication; it takes the form of robots creating noises with a hammer which mem- bers of an exhibition audience and other robots can per- ceive and react to. They moreover stress from an enactivist perspective that the robots’ actions provide “a window on the agents’ viewpoint” (Gemeinboeck and Saunders, 2013), thus possibly facilitating more introspection. Wicke and Veale refer to work outside CC to emphasise that, “when identification with the [story]teller is the goal, the physical presence of a moving body with a human shape makes all the difference” (Wicke and Veale, 2018), hence stressing the effect of organismoid (humanoid) embodiment. Moreover, they hypothesise that a “listener that can identify with the storyteller is better positioned to empathize with the story that the teller wants to convey, especially when that story is crafted from the life experiences of the listeners them- selves”. This reliance on life experience resonates with Colton, Pease and Saunder’s (2018) previously mentioned factors to improve the authenticity of CC systems. Intentionality & Autonomy While organismic embodi- ment is rarely addressed in the literature and typically only through human embodiment (Kirsh, 2011; Alexandre, 2020; Loesel, Mirowski, and Mathewson, 2020), Guckelsberger et al. (2016) highlight that organismic embodiment realised in artificial systems might play a central role in future CC re- search. Their theoretical investigation sets out by consider- ing non-artistic creativity in simple computational systems through the lens of autopoietic enactivism (Maturana and Varela, 1987) as adopted in the theory of enactive AI (Froese and Ziemke, 2009). The latter theory holds that machines with organismic embodiment can, similar to living beings, realise an intrinsic purpose by maintaining the precarious existence induced by this form of embodiment. It more- over claims that this intrinsic purpose can ground intentional agency. Guckelsberger et al. (2016) consider more specifi- cally when organismic embodiment can ground intentional creative agency, realising genuine creative autonomy (Jen- nings, 2010). They argue that a machine grounds value and novelty in creative activity through the maintenance of their precarious identity, based on acts of self-production and adaptation against entropic forces. The claim that in- tentional creative agency is contingent on organismic em- bodiment allows for additional, radical statements. Guck- elsberger et al. (2016) argue via Dreyfus (2007) that a CC system might have to accurately reproduce human organ- ismic embodiment to exhibit human-like creativity with intentional agency. Moreover, they introduce the concept of “embodiment distance” to put forward hypotheses on the impact of organismic embodiment on the perception of cre- ativity: “when we evaluate the creativity of non-human sys- tems with intentional agency, we are likely to misjudge value Proceedings of the 12th International Conference on Computational Creativity (ICCC ’21) ISBN: 978-989-54160-3-5 197  in their behaviour or artefacts, or hesitate to attribute any value at all, as our embodiment distance is too large” (Guck- elsberger et al., 2016). This makes the mimicking of or- ganismoid, and potentially, humanoid embodiment relevant. Colton, Pease, and Saunders (2018) extend the concept of embodiment distance to non-organismically embodied sys- tems and discuss ways to overcome it to foster the perception of creativity and authenticity in CC systems. Challenges of Embodied CC Our analysis exposed that many opportunities for embrac- ing embodiment in CC have a flip-side, the impact of which is mediated by the respective embodiment type. The identi- fied three challenges are easily overlooked, as they are often addressed separately from the corresponding opportunities. Computational & Design Costs While affording a range of opportunities, such as more natural interaction, the stimu- lation of creativity, grounding, etc. – being embedded in our physical world also puts high demands on [the] hardware, software and system engineering (Saunders et al., 2010) of physically embodied agents. Fitzgerald, Goel, and Thomaz (2017) particularly lament the increased processing costs due to the high dimensionality of robot sensors and actua- tors. Gemeinboeck and Saunders summarise that “embody- ing creative agents and embedding them in our everyday or public environment is often messier and more ambiguous than purely computational simulation” (Gemeinboeck and Saunders, 2013). Especially when tempted by the oppor- tunity to outsource computation into our physical environ- ment, these costs must be carefully weighted. Unpredictability Related, creative behaviour that emerges from the interaction of any embodied agent and their environment, especially if resulting from intrinsic motivation, is often hard to predict (Guckelsberger et al., 2016). This is more relevant in some application domains than others, with many artistic domains affording unique possibilities for playful experimentation. Across domains however, researchers must exercise particular caution when designing for interaction with people. Crucially, virtual embodiment comes with more well-defined interaction interfaces and affords more control in experiments that can be reset and afford stronger introspection. False Expectations Several authors express hope that or- ganismoid, in particular humanoid embodiment can facili- tate stronger identification, empathy and affect (e.g. Wicke and Veale, 2018). Saunders and Gemeinboeck however warn that humanoid robots can cause disappointment as they “elicit human investment based on superficial and often false social cues” (Saunders and Gemeinboeck, 2018). Ref- erencing studies of human-robot interaction (Dautenhahn, 2013), they particularly highlight the risk of shaping hu- man expectations in a robot’s social capabilities based on appearance alone. As a workaround, they suggest focus- ing research efforts on non-anthropomorphic robots, and on generating embodied natural movement as means of identifi- cation, instead of a similar form or sensorimotor equipment. Discussion of Embodiment Assessment We faced several challenges in assessing the type of embodi- ment in the selected papers. We briefly elaborate on embod- iment (i) classification challenges, (ii) biases, (iii) “under-” and (iv) “over-attributions”, and (v) typology limitations. The classification (i) of embodiment was complicated by unspecific descriptions especially in theoretical papers. Moreover, some papers related to several systems at once, or exclusively addressed human rather than machine em- bodiment. Sometimes the lack of specifics did not allow us to gain insights into the use of embodiment in CC. Schor- lemmer et al. (2014) for instance only appeal to embodied cognition in a side-note. Affected contributions are listed in Table 1 without an assessment of their embodiment type. Biased views of embodiment (ii) were expressed when authors only explicitly recognised embodiment in humans, e.g. in the case of a virtual system affecting the embodiment of a human user, or when discussing the challenges of mod- elling human movement. An example of this can be seen in Schubert and Mombaur (2013), who attempt to capture the movement dynamics of human painters. Some authors “under-attributed” (iii) embodiment in that they leveraged a specific embodiment without explicitly re- ferring to its type or properties. We observed this particu- larly often for organismoid embodiment, e.g. in Saunders et al.’s (2010; 2013) description of the Curious Whispers robot society. Here, the robots’ bug-like, organismoid embodi- ment is not discussed explicitly, although it may have a spe- cific effect on the perception by a human interaction partner. Closely related, other authors “over-attributed” (iv) em- bodiment in that they explicitly appealed to a certain type of embodiment without fully implementing or discussing its required properties. For example de Melo and Gratch (2010) describe an experiment performed on “virtual humans”, but they are used as passive mannequins to shine light on. Given that a simulated environment to perturb and be perturbed is absent, we cannot even attest structural coupling. Again, this was particularly evident for organismoid embodiment. Limitations of our typology (v) made it challenging to identify specific embodiment types, in particular historical embodiment: several systems presented memory or learn- ing capabilities, but this learning did not necessarily hap- pen during the systems’ lifetime. This launched a discussion amongst the authors on whether e.g. the existence of learn- ing hardware is indicative of historical embodiment. We eventually agreed that historical embodiment is independent of a lifetime criterion or learning, and present if the system’s past structural coupling has an effect on its future coupling, e.g. when a past perception triggers a change in a virtual or physical sensor, e.g. affecting a camera’s angle, thus in- fluencing future perceptions. Organismoid embodiment was also difficult to assess, as the proximity of a system in shape and sensorimotor equipment to living beings is a continuum. Given the focus of many systems on humanoid embodiment, it would have been helpful to consider it a separate type. Some types of embodiment turned out to be worse at dif- ferentiating and distinguishing existing work and insights than others, but this does not necessarily disqualify them from future use. Structural coupling for instance is the most Proceedings of the 12th International Conference on Computational Creativity (ICCC ’21) ISBN: 978-989-54160-3-5 198  inclusive type of embodiment and applied to almost all re- viewed systems, in particular to any physical object; yet, it proved valuable for sanity-checking whether a simulated system can be considered embodied at all. Organismic em- bodiment in contrast was the most exclusive type. We think it should be included nonetheless, given that existing theo- retical work (Guckelsberger, Salge, and Colton, 2017) as- signs it a major role in future CC research, e.g. on creative autonomy. As for any other type, the current lack of dif- ferentiating power might indicate an under-appreciation in existing work, rather than a weakness of the type itself. Directives for Embodied CC Research Our analysis supports the hypothesis which motivated this systematic review in the first place: creativity in artificial systems, and how it is perceived, is affected by embodiment. If we assume this hypothesis, then furthering our insights into embodied CC will play a critical role in advancing the goals of CC more generally. However, existing research is highly fragmented and ambiguous, lacks generalising empir- ical results, and rarely trades-off the opportunities and chal- lenges of a certain type of embodiment. We translate find- ings from our review on RQ1, RQ3 and RQ2 into three di- rectives as pillars of a future embodied CC research agenda. Clarify embodiment: Our review highlighted that many opportunities and challenges of embodied CC are exclusive to a specific type of embodiment. However, we experienced serious difficulties in assessing the specific type of embodi- ment embraced in existing work, as documented in the cor- responding papers. In order to establish an efficient, unam- biguous, verbal and written scientific discourse on embodied CC, we urge researchers to always clarify what specific em- bodiment they appeal to in a particular theoretical or applied project. To this end, adopting definitions from typologies as presented here may serve as a shortcut, foster comparison, and alleviate the “under-” or “over-attribution” of a certain embodiment. However, any typology can be contested and lacks detail; references to specific embodiment types should thus always be complemented with extensive descriptions. To counteract the fragmentation of research output, we recommend to reference embodiment-related work exten- sively. This review, albeit non-exhaustive, might serve as a starting point to identify relevant work. Conduct empirical studies: Our review moreover uncov- ered that no existing CC study produced generalising, em- pirical insights about the effect of embodiment on (i) an artificial system’s creativity, and (ii) how its creativity is perceived by others, including humans. Researchers either make assumptions on this relationship, or draw on exist- ing empirical findings from other fields that may not easily translate to computational systems (e.g. Brown, 2016). We recommend to conduct qualitative and quantitative empiri- cal studies on the impact of a specific embodiment, treated as independent variable, on (the perception of) creativity. Informed by our analysis of which embodiments differen- tiated existing work most, we recommend that initial empir- ical studies should investigate virtual and physical, or hu- manoid and other variations of organismoid embodiment, as values of the independent variable. In evaluating the perception of creativity (ii) as depen- dent variable, experimenters must eliminate or weight for creative ability (i). Vice versa, in evaluating creative ability (i), they must avoid bias by the perception of creativity (ii), e.g. by employing objective measures of creativity (Ritchie, 2007). When employing subjective measures, researchers should consider concepts introduced in the literature such as the embodiment gap (Guckelsberger, Salge, and Colton, 2017; Colton, Pease, and Saunders, 2018) between the sys- tem and its evaluator(s) as mediating variable. Trade-off opportunities and challenges: We encourage researchers to eventually make use of these empirical in- sights for the design of systems that reliably leverage a cer- tain embodiment as a means to an end, e.g. to accurately model creative processes that emerge from sensorimotor in- teraction in a specific domain. Crucially though, our re- view showed that most embodiment-related opportunities come with a challenging flip-side. In order to avoid an un- favourable trade-off, we recommend researchers to always inform their choice of a particular embodiment not only by its opportunities, but also by the corresponding challenges. Conclusions and Future Work Motivated by the potential impact of embodiment on cre- ativity and its perception in artificial systems, we set out to map the present landscape of embodied computational cre- ativity (CC), and offered directions for future research in this area. To this end, we conducted a systematic review of papers presented at the International Conference on Com- putational Creativity (ICCC) that explicitly discuss embod- iment. To counteract ambiguity in the concept’s use, we adopted a well-established embodiment typology, and ex- tended it based on the recent scientific debate. We found that most existing work can be differentiated by its focus on virtual vs. physical, and, more fine-granularly, on humanoid vs. non-humanoid, organismoid embodiment. Moreover, we showed that each type comes with its unique opportunities and challenges as flip-sides of the same coin. Overall, we identified nine opportunities, e.g. the outsourcing of compu- tation or support for more natural interaction, and three chal- lenges, e.g. unpredictability and the shaping of false expec- tations, from existing studies. We identified several short- comings of existing work that likely hinder progress on em- bodied CC research, most prominently ambiguity in the use of the embodiment concept and a lack of dedicated empiri- cal research. We leverage these insights in our final contri- bution: three directives to advance embodied CC research. While we chose the scope of this study to provide a rea- sonably unbiased, big-picture view of embodied CC, future work should be dedicated to incorporating additional, rele- vant references. How to meaningfully constrain the scope is an open question, as relevant work can not only be found in be found in CC books, but also related fields such as videogame AI, robotics, design...