From Disembodiment to Embodiment in Artificial Intelligence and Psychology - Parallels in Thinking

This paper briefly traces how both Computer Science and Psychology grew-in tandem-to share similar notions of Embodied Cognition. It concludes an analysis of how the two fields are uniting in their work on Embodied Artificial Intelligence, and future challenges for the field.


Artificial Intelligence and Embodied Cognition
The Turing computer became the central model for cognitive science during the 1950s and 1960s (Sprevak, 2017), and was reinforced by Putman's Computational Theory of Mind (CMT) (1961).Not long after, Jerry Fodor (1975Fodor ( , 1983) advocated for mental processes consisting of computational operations that were symbolic representations within the brain.Ultimately, his ideas were dominated largely by psychological nativism, and his legacy became synonymous with the modularity of the brain and domain specificity (Fodor, 1975(Fodor, , 1983(Fodor, , 1990)).During this time, developments in linguistics (see Chomsky, 1965), in cognitive science and in artificial intelligence joined forces to usher in the cognitive revolution (Smelser & Baltes, 2001).
In 1980s, John Hopfield and David Rumelhart promoted deep learning techniques that allowed computers to learn using experience/input (Hopfield, 1984;Rockwell, 2017).Deep learning and the idea of parallel distributed processing (Rumelhart & McClelland, 1987;Rumelhart & Ortonty, 1977;Rumelhart et al., 1986) became the basis for questioning traditional models of cognition (disembodied) in favor of models of embodied cognition.Similarly, deep learning solved AI's central problem of representation (Hopfield, 1984) and combated the failings of representational methods (Duffy & Joue, 2000).Specifically, this meant that learning could emerge from multilayered distributed processing models (through rapid accelerations and decelerations) and exhibit stage-like effects, suggesting that emergence Similarly, Brooks (1991Brooks ( , 1999)), also influenced by J.J. Gibson's work, challenged the view that intelligence was the result of bigger brains and larger central processors to encode elaborate representations.Rather, he argued that millions of years of evolution within the environment was responsible for the complexity of human responding.He went on to champion the position that robots could be designed in a bottom-up manner, with sensors and movement through their environment to navigate the world (see Morgan, 2018).

Psychology and Embodied Cognition
Much like AI, Psychology in the 1960s, ushered in the 'cognitive revolution,' challenging behaviourist theory and endeavouring to discover how the brain functions and learns.The mind was now thought of as running on the hardware of the brain (Dennett, 2014).Central to these ideas was that of "representation", in which internal symbols "stood in" for the environment outside the body of the organism.Moreover, such mental representations within the brain were usually thought of as abstractions of the original information.This computer analogy, an updated version of the age-old Cartesian model, continued to promote the disembodied view of brain-bound cognition (see Fugate et al., 2019;Macrine & Fugate, 2020, 2021, 2022;Thompson & Cosmelli, 2011).The body was seen as a "passive" observer of the brain, and necessary only in the execution of motor actions (Macrine & Fugate, 2020).
Beginning in the late 1970s, however, this view began to change, but not without input from some famous predecessors.John Dewey (from works between 1925Dewey (from works between -1953)), echoing William James (1890), suggested that higher-order cognitive functions were adaptations generated by interactions with the world.J.J. Gibson's "ecological theory" (1979) married both phenomenological (i.e., the subjective experience) and naturalistic perspectives.Gibson (1979) argued that perception was direct, and the environment meaningful.He advocated against representation and argued rather that organisms make sense of their world through direct interaction with it.For Gibson, affordances allowed an organism to interact in specific "smart" ways with its world: An organism and its adaptive environment have a "goodness of fit", and bodies, as well as their perceptual and motor functions, did not evolve independently of one another.Importantly, this co-adaptation ensured little need for an internal executive processing system or representation of information, and the cognitive burden was offloaded to the environment (see also Anderson, 2003;Clark, 1999, Nöe, 2009;Wilson, 2002).Both James and Dewey rejected the "rational psychology" drawn from Cartesian Dualism.
In 1991, Varela, Thompson, and Rosch introduced the world to embodied action and formalized what many considered the original view of Embodied Cognition (EC).According to their theory: 1) Cognition is dependent on experiences that come from having a particular body design (and differences in design among animals); 2) Experiences of the body are embedded in a more encompassing biological, psychological, and cultural context.Specifically, Varela et al. (1991) write: "The perceiver, with its body and sensory and motor systems, will determine what is available to that organism, such that as an organism moves through the world, its motion will produce opportunities for new perceptions and new perception will reveal opportunities for new actions!"(p.173).
Today, there are many versions of EC.Most reject the computational metaphors of the brain in favour of ones that stress the intertwined nature of sensory and perceptual information with action (Brooks, 1990;Grosso et al., 1995;Pfeifer, 1999).EC argues that an organism's cognition is grounded in both perception and action, such that perception is for action, and deeply dependent upon features of the organism's physical body (e.g., Barsalou, 1999Barsalou, , 2008; see also Clark, 2008;Golonka & Wilson, 2013;Lakoff & Johnson, 1999;Pfeifer & Bongard, 2007;Shapiro, 2011;Willems Francken, 2012).In addition, most theories suggest that bodily sensations interact with physical environments, and -in the case of humans-with their social and cultural contexts (Barsalou, 2010;Núñez & Lakoff, 2005;Wilson, 2002).
While different theories of EC differ in the extent and nature of such representations, central to most of them is the idea of amodal abstraction, in which the latter refers specifically to the representation outside of the sensory-motor areas.To help explain how such activation is used conceptually, Barsalou (1999) developed the Perceptual Symbols Systems (PSS).Accordingly, thinking about action will evoke the same visual stimuli, motor movement, and tactile sensations, etc., that occur during the action itself (Barsalou, 1999(Barsalou, , 2003(Barsalou, , 2008)).The experience can be later recreated (through simulation) without action (i.e., when just thinking).
During the last few decades, theories of EC have created a paradigm shift in cognitive psychology from a reactive model of the brain to a generative and predictive processing model (McNerney, 2011).Today, the mind is no longer conceived of as a set of logical/abstract functions, but as a biological system rooted in bodily experience and interconnected with bodily action with other individuals and the environment (Adenzato & Garbarini, 2012).From this perspective, action and representation are no longer interpreted in terms of the classic physical-mental state dichotomy but are closely interconnected (Marmeleira & Duarte Santos, 2019).Embodied cognition frameworks include embodied (body-based) thinking; embedded (situated) thinking, extended (in the environment) thinking, and enacted (dynamic interaction) thinking (Gallagher's 4Es, as cited in Rowlands, 2010; Newen et al., 2018).Expanding on these ideas, radical enactivism suggests that the body is not just a means but also "an end of having a cognitive system" (di Paolo & De Jaegher, 2017; see also Hutto & Abrahamson, 2022), Thus, cognition is less a matter of mental reasoning and abstract problem solving, than a matter of adaptive self-regulation (di Paolo & De Jaegher, 2017) and enacting in an environment (Hutto & Abrahamson, 2022).Accordingly, the exact way an organism's body performs certain interactions with the environment leads to perceptual recognition (Kumar, 2008).

Parallels Between Psychology and AI in Embodied Cognition
The recent developments in cognitive science and neuroscience have contributed to "predictive processing" theories of mind and cognitive function (Clark, 2013;Friston, 2011;Hohwy, 2013).Such approaches are similar in that they emphasize top-down, error-minimizing predictions as the key locus of information processing, as opposed to feed-forward feature recognition, common in more classical computational theories (Allen & Friston, 2018).This embodied turn in AI indicates that the body shapes thinking, and thinking is bound and realized through our embodiment (Pfeifer & Bongard, 2007).Therefore, when thinking is divorced from the body, critical elements are lost.For example, intelligent behaviors can be rapidly learned by agents whose morphologies are adapted to the affordances in the environment (Brooks, 1991;Bongard, 2014;Gupta et al., 2021;Pfeifer & Scheier, 2001).At a more proximal level, this involves consideration of what sensors an agent can use, how its actuators operate, and how its perceptions and the dynamics of its movements are affected by its physical body (Iida & Nuraman, 2016).Pfeifer and colleagues (2007) wrote that the material properties of an organism's musculoskeletal system, the sensor morphology, and the interaction with the environment, all provide an "alternative avenue for tackling the challenges faced by robotics.The tasks performed by the controller in the classical approach are now partially taken over by morphology and materials in a process of selforganization" (p.1098).Thus, true AI will only be achieved when robots can connect sensory and motor skills through a body (see Brooks,1991;Pfeifer, 2001Pfeifer, , 2006)).
The direction of AI has changed from a purely computational discipline into a highly transdisciplinary system based on the paradigm shift of EC. "With the fundamental paradigm shift from a computational to an embodied perspective, the nature of research topics, the theoretical and technical issues, and the disciplines involved have changed dramatically, or in other words, the "landscape" has changed completely" (Pfeifer & Iida 2004, p.1)This embodied focus has led to the creation of the field of Embodied Artificial Intelligence (EAI).EAI rejects computationalism and its limitations, instead focusing on intelligent systems established through the interaction of an agent with the environment through dynamical predictive processing of sensory-motor activity (Bermudez, 2021;Brooks, 1990;Pfeifer, 1999).The idea here is that intelligence emerges, rather than being learned exclusively from datasets of text and images (Duan et al., 2022;Gupta et al., 2021).Supporting this thesis, Hinton (2013) argued that EC continues to have an increasingly influential perspective in the user-experience design fields and stands to fundamentally change the way we think about and design human-computer interaction.

Summary and Future Directions: EAI and Evolutionary Robotics
More recently, EAI researchers are looking at human behavior and embodiment, to try to give both hard and virtual agents life-like capabilities (Floreano, et al., 2014, p.1). Pfeifer and Iida (2004) argue that one of the major challenges in EAI is an ecological balance-for example, in both hard and virtual agents there should be a match in the complexity of the sensory, motor, and neural control systems.In the case of hard robotics, they are built of hard materials and electrical motors and lack a sophisticated sensory-motor system and still require an enormous amount of computation (Pfeifer & Iida, 2004).
Our discussion would not be complete without a quick mention of the field of Evolutionary robotics (ER).Within the last ten years, ER as a field provides the tools for simulating embodied agents in a realistic environment to study evolutionary constraints on behaviors as they evolved, namely with respect to the interactions of an organism's morphology with the environment over time.Thus, ER addresses successful adaptations as pre-programmed mutations which lead to the end goal -discarding and replacing failed designs with mutations and combinations for better designs.The idea is to generate an accurate "design landscape" through the application of high-throughput techniques and "tightly coupled digital-experimental systems" (Howard et al., 2022, p.1).Whereas early evolutionary robotics were mainly virtual, today with 3D printing, robotic "bodies" can handle more advanced body-based processing.Several issues still exist within the field, however, including the problem of evolving in "realtime" to a rapid, ever-changing environment to drive learning (Howard et al., 2022).Alongside this push there needs to be a focus on task-based interactive questions, in which an embodied agent performs specific tasks in the real world to unlock new insights (Duan et al., 2022).For instance, an agent might have to boil an egg to learn how long it takes to do so.From this, the action learned would explicate the knowledge to develop generalized task planning, which is crucial to developing general all-purpose AI simulations in the real world (Duan et al., 2022).Understanding the limits of what is possible for an ER is a critical question that remains to be addressed.For example, understanding exactly how morphology and sensory modality impact the ability to learn (i.e., the tradeoffs between power, mass, computing) are still not fully understood, and how these create inductive biases in learning, especially because many robots are built for one environment and fail when presented to a new environment (Roy et al., 2021).