(Note: This post possibly assumes a bit of familiarity with ideas around Gestalt phenomenology, and the relationship of this to ethnomethodology. Quéré [PDF] has a useful description: "When we perceive a Gestalt, we grasp more than mere outlines and shapes; we perceive an ordered and differentiated whole, or an arrangement of parts. As a form, a Gestalt is a unity of order: it holds together and orders the differentiated parts of a whole.")
I am part of a migratory flow of (broadly put) human-centred technology researchers that is pushing into robotics, human-robot interaction, and relatedly "the new AI" (aka "generative AI" technologies). As part of this shift I frequently get sight of people invoking ideas about "embodiment". Interest in embodiment and embodiment-family concepts like "tangible" and "physical" computing entered into HCI research to the best of my knowledge around the late 1990s, perhaps most prominently with work by Dourish (e.g., 2001's Where the Action Is). It's now just taken to be part of the furniture in HCI, one perspective amongst many (see Oulasvirta & Hornbæk's 2017 paper What Is Interaction?). But with the increased popularity of (and vast capital flowing into) both robots and the intersection of robotics and "AI", coupled with the intellectual capture of large swathes of academic research interests that comes with this, it feels to me that "embodiment" is having another moment in the research discourse. (Or maybe I am just starting to notice it more?)
If you are an HRI researcher, I realise what might be coming to mind here. To be clear I'm absolutely certain HRI researchers and robot designers of all kinds have, in reality, long been occupied by embodiment-family concepts, so I am not claiming community ignorance. One obvious hinge-point is Brooks's perception and action approach in robotics [PDF] from the 1990s that marked a critical shift away from focussing on inner representations, creating greater emphasis on researching robots as complete systems in-the-world. It could even be that the "embodiment" factor is an important reason why HRI researchers seem to really like robots in the first place - their physical presence, materiality and ability to actuate (note, as distinct from act), as fascinating "autonomous beings" made manifest. As Fantasia et al. state [PDF] "robot embodiment" is "the defining feature which separates HRI from [HCI]".
I also see a growth in embodiment-family concepts being leveraged within "AI" cognate research, partly but not exclusively at the human-centred end and particularly in how "AI" is spoken of in the same sentence as robots. (Note, my use of AI + quotation marks = reminder of the problematics of this term [PDF].) For instance, I see a rise in popularity of "embodied AI", "embodied intelligence" or "physical AI", notions that frequently intersect with robots and HRI. Again, this is not to say that embodiment-family concepts are actually novel to "AI" discourse at all. One only has to turn back to Dreyfus's critiques of the 1960s and 70s to locate concerns around embodiment and its significance (Dreyfus 1992, p.235). Dreyfus argues that being embodied is in fact critical to true "intelligence". He also goes further, pointing out that embodiment itself is saturated by part-whole relationships (more on that later). By this he means that - following the influence of Merleau-Ponty - human experience of embodiment exists within a reflexively constituted Gestalt, where past experiences with the specific determinate objects (aka Husserl's "inner horizon") and the whole situation ("outer horizon") are brought together to make sense of things. Dreyfus muses about then-current attempts by Minsky and others to embody "AI" in experiments with robot arms and the prospect of this Gestalt approach continuing to pose fundamental problems for computation in spite of the move to more "embodied AI" (my use of the phrase not Dreyfus's). (Footnote: Dreyfus gives examples of the Gestalt issue with what we are pushed to call "embodied activities", such as drinking a glass of "water" that turns out to be milk, or the relationship between a single note and a whole musical piece. The chapter in his 1992 revised edition is still worth reading in full.)
But that all said - the point of this post is to tell you why I often feel conflicted on embodiment-family concepts, primarily in their use as analytic terms within the kinds of human-centred research that crosses into territories of robots and "AI". This is motived by my realisation at some point that I was probably avoiding using "embodiment" - or at least felt faintly ridiculous when using it. To satisfy my own curiousity I looked back on a recent CSCW paper "On human-robot collaboration" [PDF], which is entirely about bodies in action and how those bodies fit alongside a public delivery robot service - but found no mention of "embodiment" or anything like that.
So it's time to reflect on why I feel like this. What exactly is troublesome to me about embodiment?
To begin with, embodiment as a concept has a tendency to both construct and then narrow analytic windows. By this I mean that as soon as "embodied X" or reference to "an embodiment" is invoked as an analytic category it can arrive with the implication that there are (or can be) distinct domains of "embodied" and "not-embodied" things, actions, etc. It can imply even that there is the possibility of a dimensional scale of embodiment. You can see this in discussions within HRI that compare "embodied" (e.g., physically present robot) with "disembodied" (e.g., videoed robot "stimulus", or perhaps a simulated one) and seek to work out whether this matters or not for coming up with HRI findings. I doubt this is so much a serious engagement with, say, phenomenology. Instead I suspect such interests are concerned with generating research findings that can be arrived at with appropriate efficiencies of time and money. (With the construction of such embodied-disembodied comparisons I hear overtones of Hollan and Stornetta's arguments about the mistakes of remote communications design imitating face-to-face [PDF]. I wonder if anyone has considered the analogy here with HRI's attitudes towards embodied-disembodied comparisons?)
A more essential, pragmatic reason why such categorisations (e.g., "embodied X") might be put into play is because researchers want to draw attention to something about the physical, material stuff in front of us. This can be a way of creating some kind of distinction - for instance, in HCI it might mean effectively saying "not screens", or at least meaning that what is happening "on screens" is not the main interest. Naturally, emphasising the "embodied facets" of HCI is one way of trying to carve this out, perhaps trying to pick out the role embodiments play in interaction with digital interfaces. In the earliest cases, such as that of Dourish's work, there was a clear strategic and paradigmatic sense to doing this: specifically trying to effect a shift in understanding within a research field (of HCI) that the (certain kinds of) physical stuff matters. This seems justifiable even if it introduces some analytic problems that we can ignore for a bit. But such shifts have a tendency to become parodies when we treat them as analytic developments. So, the strategic move instead becomes "the embodied turn": a specific new paradigm of research where we centre certain types of physicality (i.e., not-screens) and certain attributes like human bodies (and with this, individuals). A problem is that this rides roughshod over, say, what Introna and Ilharco call the "screenness" of screens [PDF] - a kind of HCI path that never really was explored much in my view, perhaps because non-screen based interaction is just less boring (probably true, given how a 2016 paper we wrote on this [PDF] seems to have disappeared without trace). For research fields like HCI and HRI which lack much in the way of agreement about core phenomena, epistemology, and even method, concepts like "embodiment" can prove useful as a therapeutic, catholic anchor [PDF]. So, while I appreciate the distinction afforded by deploying embodiment-affiliated concepts like "physical AI" as ways of communicating research to specific research audiences, we really should not take it much more seriously than that - i.e., as a matter of research branding.
Use of embodiment-family concepts is also, I think, subtley reflective of researchers' own phenomenal experiences of technology. This might be in navigating the seeming ephemera of their personal engagements with social media, mediated communications, mobile interfaces or whatever, and the sense this brings of the deep ungraspable that permeates encounters with these technologies and their supportive (hidden) infrastructures. This ungraspability, ephemerality, etc. is absolutely not true, of course, and it is clearly the case that all technologies are necessarily caught up in the grist of everyday action which always includes "bodies", "scenes", etc. But seeing that this is the case can be difficult because we are members (of social milieu) first, and reseachers second. Once again however, I have to concede that embodiment-family concepts can still act as a useful aggregator for the matter of distinguising experiences.
Reiterating a previous point, an obvious and unavoidable analytic problem with embodiment is that all things we study in HRI, HCI and "human-AI interaction" happen some-where, some-time. It is much easier to pretend that it doesn't, and then add "embodiment" in at some point later as a special new facet of study. Uncharitably, this follows the trajectory of cognitive science - with exclusive focus on the "inner" (mind) being found wanting, increased rejection of representationalism, and thus various expansion operations have been attempted to pull together the outer (the world "out there") with the inner (the mind). These attempts travel under names like distributed cognition, embodied cognition, embedded cognition, and so on (see Hutchinson's paper on 4E cognition). Embodiment-family concepts are being used here as a retort to restrictive kinds of representationalism native to cognitive science; the retort being that the inner cannot be made sense of without the outer. This feels also in line with Dreyfus's arguments.
Instead of all this confusion, my proposal is that the key question is not whether or exactly how this or that thing, activity, or process is or can be found to be "embodied". But rather the question is whether features we might group together with concepts like "physicality", "embodiment", etc. are themselves relevant to socially ordered activities and projects at hand for people in the scenes as we find them. The matter of the relevancy of embodiment is made in situ, and must be found in situ as well.
Physical pain, for instance, can radically transform everything such that "bodily experience of pain" is suddenly and persistently pertinent to all manner of our goings on in the world. We did not shift from a state of "disembodied" to "embodied". Rather, the praxeological Gestalt (cf. Hutchinson [PDF]) or "fabric" of everyday life has shifted, gotten rucked up somehow. The bit where we trip over it is the bit you are calling "embodied". Embodiment conceptually is therefore not necessarily special or virtuous even if we wish it would be. Instead, "embodiment" is just the name for a mundane and often "invisible" (not presently relevant) aspect of all things - inseparable from their formation through everyday action. As a potential analysis, "embodiment" is always sitting there in the background, often unattended to. This makes it easy to then claim embodiment's relevance, analytically, to anything, even if this countermands what people in an unfolding scene treat as important.
The other vital aspect here is derived from my ethnomethodological (EM) reading of "embodiment". That there is a tendency for the concept of embodiment and its use - like some phenomenological readings - to mean centring individuals. I think Dreyfus is somewhat guilty of this. Speaking of this or that action being "embodied" is has the potential to reify the "fabric" I described above - i.e., praxeological Gestalts and their contextures (i.e., including their horizons of anticipation and storied character, their universal indexicality, etc.). Reifying is then collapsing "embodiment" into an attribute of a person. The individualism of embodiment - possibly a throwback to its potential cognitive origins, or at the very least cognivist resonances - therefore means we are missing how these are originally social phenomena - or to put it another way, matters of collectivities - and not simply products of individuals alone that are merely aggregated somehow. This perspective is very much oriented by EM, which, in its studies is really about "establishing the nature of the internal relationship between particulars and the whole", as quoted from Hutchinson, who uses the term "Praxeological Gestalten" which nicely calls back to my comments above on Dreyfus. (Note, I have been / will be using the terms like "scenes" often as shorthands for Gestalts.)
I am now stacking further problems up here, because the analytic issue is also that the conceptual concerns of "embodiment" are typically analysts' concerns, not those of people (aka Garfinkel and Sacks's "members") in the scene who are tripping over robots or getting "advice" from "AI". People can or may be concerned with many particular, evolving matters that are all deemed to fall within the remit of "embodiment", such as the mechanics of specific body parts, the visual appearance of things, sensations felt, spatial configurations, touch, aural qualities, etc. These cannot be collected together as simply "matters of embodiment" that are left unalloyed to the sense they derive from their placement or storied character within ongoing scenes. Consider, for instance, the concision and categorial sensitivity with which nurses and elderly patients jointly work through attributions of gender to a humanoid robot administering a survey to them in Stommel et al.; features related to the visual appearance of the robot, for instance, are made to buttress such attributions amongst many other "embodied" elements (see the paper for more). Or, returning to the pain example, Ryle pointed to the incommensurate features of pleasure and pain [PDF], yet they are typically bundled blandly into "embodiment". As ever, we should be directed by the ways that people on scene articulate and are sensitive to the complexities of these quite different categories of phenomena.
Because of the lack of analytic depth prevalent in discussions in human-oriented "AI" research, ideas about "physical AI" or "embodied AI" often just amount to "we used a robot". This formulation is a result of the conceptual morass that comes with setting up shop with "artificial intelligence". "Intelligence" is traditionally tied to something "in the head". So you simply put it into the slot waiting in "the body" - hence "embodied AI". In HRI it might go further and be about measuring the "amount" or "level" of embodiment, or perhaps even how robot cognition and embodiments integrate with the cognition and embodiments of human "collaborators". But it is all somewhat limited conceptually seemingly no matter how many "think pieces" are written about "AI".
To conclude, I have been trying to describe a different way to think about analysis of technology like robots and whatever "embodied AI" might mean. Rather than fuss about the implications of "embodiment" we can instead look to working out how robots or "AI" systems of one type or another must necessarily become situated within what Garfinkel calls "immortal ordinary society" - that is, the objectivity and persistence (in our absence) of social, worldly life that consistently and repeatedly reproduces orderly scenes again and again, with or without us. One of the problems with this alternative is that its discovery must be practiced to be understood. It cannot be treated as a theoretical matter. You have to instead see the sense of it as a practical engagement with everyday scenes as we find them that also turns upon the capabilities of oneself (e.g., a researcher) also as a member of these scenes. To me this is one of the things I have been forcibly reminded of in our work on studying robots in public. The challenge is that this can be a real hurdle for researchers trained to look for, recognise and then generate constructive relationships between "theory" (e.g., embodiment) and "practice" (e.g., technologies' incarnate necessities). My point is that one cannot start a journey from a pre-existing analytic point of studying "embodied AI" or studying "robot embodiments". Instead you have to reach those concerns as the work of the studies themselves - which is why I highlighted Stommel et al.'s paper on robot gender attribution in particular as a real exemplar of this approach and sensibility.
(Updated 11th Sept 2026.)