Current AI motion systems are increasingly good at representing what a body looks like and where its joints are, yet those achievements should not be mistaken for a body schema: the largely pre-reflective organisation that makes situated action possible. The distinction between body image and body schema gives motion AI a precise architectural critique. A generated skeleton, camera-registered rig or pose embedding may model the body as an object while still lacking the action-oriented, adaptive organisation through which a body moves.

This analysis develops that distinction and applies it to motion representation, video-to-humanoid control and body-to-AI interaction. The connection is interpretive rather than a claim that phenomenological concepts map directly onto neural modules. Its value is diagnostic: it reveals which bodily capacities current systems externalise, approximate or leave unresolved.

Body image and body schema answer different questions

Shaun Gallagher's formulation distinguishes a body image—perceptions, beliefs and attitudes directed toward one's own body—from a body schema, a system of sensorimotor capacities that operates without requiring the body to become an explicit object of attention (Gallagher, 2005). Frédérique de Vignemont's review shows that the terminology has varied across disciplines, but the conceptual pressure remains: representing a body and organising action through a body are not the same function (de Vignemont, 2010).

The distinction is easiest to feel in ordinary action. Reaching for a cup does not usually require calculating the shoulder, elbow and wrist as visible points. The body is not absent; it is operative. Proprioception, balance, learned habit, environmental affordances and ongoing correction participate before a reflective picture of “my body” is assembled.

Maurice Merleau-Ponty's account of bodily intentionality deepens this point. The lived body is not simply an object located in geometric space. It is the perspective through which a meaningful field of possible action appears (Merleau-Ponty, 2012). A handle is graspable, a passage is narrow, a rhythm invites continuation. These are relations between a particular embodied capacity and a situation.

Body schema should not be romanticised as an infallible inner intelligence. It is learned, plastic and vulnerable; it can be altered by injury, tools, training and social conditions. Nor is body image merely visual. Belief, affect and cultural judgement participate. The point is functional differentiation, not a clean anatomical partition.

Motion AI often begins from a third-person body image

Most motion datasets describe a body from outside: joint coordinates, rotations, meshes, video frames or action labels. These representations are indispensable. They allow comparison across sequences and support retargeting, generation and visualisation. But they make the body available primarily as an observable object.

A latent motion representation compresses those observations into features useful for reconstruction, classification or generation. Even when its geometry preserves temporal structure, as SemTAC attempts through local temporal contrasts, the representation is learned from recorded sequences and an objective chosen by designers (Rodrigues & Putze, 2026). It may retain how movement unfolds without possessing the first-person regulatory relation that produced it.

The difference matters because similar visible trajectories can arise from different organisations. A reach made with stable support and a reach made while recovering balance may pass through comparable keypoints. A dancer can reproduce a shape while changing breath, attention or muscular recruitment. Conversely, bodies with different proportions or mobility may organise equivalent intentions through visibly different trajectories.

The Lab's discussion of movement between frames established that temporal transition is information. The body-schema distinction adds that transition is still only part of action organisation. What is sensed as possible, costly, risky or inviting also matters.

Retargeting exposes the difference between appearance and enactment

Motion retargeting transfers a recorded movement to another skeleton or robot. The task makes the conceptual distinction concrete. A successful visual match can be physically infeasible; a stable execution can erase the movement's amplitude, lean or dynamic emphasis.

Bao and colleagues explicitly report that lower average positional tracking error does not guarantee preservation of expressive characteristics. Their residual-control framework adds objectives for amplitude, energy, body expansion, body lean and pelvis height (Bao et al., 2026). The work remains a preprint, and the five characteristics are proxies rather than expression itself. Still, its diagnosis is important: coordinate accuracy is not sufficient evidence that a receiving body has organised the action comparably.

ViT-H takes another step by using a humanoid tracker's outcomes as feedback for the motion generator (He et al., 2026). This closes a loop that conventional pipelines leave open. The generated reference is shaped by whether the robot can execute it. Architecturally, that resembles one aspect of body-schema function: action possibilities constrain representation.

Yet the resemblance has limits. A tracker's reward and success criteria are designed externally. The robot's “feedback” is not equivalent to lived proprioception, pain, fatigue or felt confidence. Optimising executability may also favour conservative motion and suppress expressive risk. A system can become more action-aware while remaining indifferent to experience.

Camera registration gives the body a scene, not yet a situation

Joints, Camera, Action! co-generates video, arbitrary-rig motion and a camera estimate, allowing the movement to be placed back into a 3D scene (Rotstein et al., 2026). This is an advance over motion that floats without a consistent viewpoint. Camera recovery establishes geometric relation among joints, frames and scene.

Phenomenologically, however, a scene is not yet a situation. A scene is a spatial arrangement available to observation. A situation includes what matters for an acting body: a surface that supports weight, an obstacle that threatens collision, another person whose timing solicits response. The same geometry can support different action possibilities depending on skill, history and condition.

This is why adding more scene labels does not automatically produce embodiment. An architecture needs feedback that changes with action and preserves the consequences of error. World models, physics simulation and embodied control can provide parts of that loop. None alone settles how relevance emerges.

What body-schema theory contributes to architecture

The distinction suggests four design requirements.

1. Represent capacity with pose

A motion state should include uncertainty, contact, support, reachability and actuator or performer-specific range—not only joint position. Capacity is dynamic: fatigue, injury, costume and floor condition can change it within a session. A model that treats one calibration as timeless turns a body into a static specification.

2. Learn through consequence

Action representations should be updated by what happens when movement is attempted. ViT-H's tracking feedback is one technical example. In a creative system, consequence can also include a performer's correction, a collision, a missed musical cue or a visual response arriving late. The expressive-fidelity audit provides one practical way to retain such evidence.

3. Preserve multiple coordinate relations

Body-centred, world-centred, camera-centred and partner-centred frames answer different questions. Flattening them into one coordinate system can hide the relation a movement expresses. A turn toward a partner is not equivalent to the same rotation in an empty world frame. Provenance must therefore include how coordinates were established and transformed.

4. Keep first-person correction in the loop

The person moving should be able to say that a visually accurate result feels wrong, and that report should not be dismissed as noise. Micro-phenomenological and soma-design methods offer disciplined approaches to experience rather than treating first-person evidence as an unstructured preference (Prpa et al., 2020). Such reports do not replace kinematic data; they identify variables the instrumentation may have omitted.

What AI architecture contributes back to somatic theory

The exchange should not run in only one direction. Computational systems force theoretical claims to become operational enough to test. If a project says it preserves “body schema,” what observable difference should follow? Does a capacity-aware representation generalise better after a change of viewpoint, tool or morphology? Can a performer adapt with fewer corrections? Which failure distinguishes a missing body relation from a poor pose estimate?

AI also makes multiplicity visible. A generated movement can be retargeted across bodies, revealing which relations survive morphology and which depend on a specific anatomy. Simulation can perturb contact, delay or field of view in controlled ways. These techniques cannot recreate lived experience, but they can sharpen questions about plasticity and adaptation.

The danger is operational capture: once a proxy can be measured, it may replace the concept. “Body schema score” would be especially misleading if it collapsed support, habit and situated possibility into one benchmark. Theory should constrain the humility of measurement as well as inspire new features.

The unresolved problem is normativity

A body schema does not merely predict the next pose. It helps organise what counts as a viable action for this body in this situation. Viability has normative force: some outcomes preserve balance, avoid harm, fulfil an intention or sustain relation.

AI systems receive such norms through objectives, demonstrations and feedback. Who sets them is therefore a central question. A robot-control objective may privilege stability over expressive extremity. A dataset may privilege motions that optical capture can see. A creative interface may reward similarity to a reference when the performer values transformation.

Chinese dance theory's shenyun (身韵), often translated incompletely as bodily rhythm or bearing, is a useful caution. Its significance is not stored in a pose alone; it is cultivated through culturally organised relations among breath, rhythm, intention and form. Converting 身韵 into a few motion descriptors may aid a specific interaction, but it cannot exhaust the practice that gives those descriptors meaning.

Normativity also complicates personalisation. A system may learn that one performer usually avoids a range and then optimise future suggestions around that history. This can be supportive when the range reflects pain or access needs, but restrictive when the performer is deliberately exploring beyond habit. A capacity model must therefore distinguish a current constraint from a permanent identity and allow the person to revise what the system assumes. Adaptation without contestability turns yesterday's body into tomorrow's limit.

Toward systems that know the limits of their body model

The immediate architectural goal should not be to claim that an AI has a body schema. It should be to build systems that distinguish their body model from the action capacities they have actually tested.

That means reporting whether a motion was merely generated, geometrically registered, simulated, physically executed or experienced by a person. It means keeping uncertainty and correction histories. It means evaluating preservation as well as task completion, and allowing multiple bodies to define viable solutions.

Body image and body schema offer a disciplined vocabulary for this hierarchy. Motion AI has made impressive progress on bodies that can be seen. Its next challenge is not simply higher-fidelity imagery. It is accountable relation among representation, capacity, consequence and situated action.

References

  • Bao, L., Zhu, H., Pan, Y., Wang, J., Yu, J., Peng, T., Jia, Y., Kanoulas, D., & Zhou, C. (2026). Beyond kinematic tracking: Preserving expressive characteristics in humanoid whole-body control. arXiv. https://arxiv.org/abs/2610.19393
  • de Vignemont, F. (2010). Body schema and body image—Pros and cons. Neuropsychologia, 48(3), 669–680. https://doi.org/10.1016/j.neuropsychologia.2009.09.022
  • Gallagher, S. (2005). How the body shapes the mind. Oxford University Press. https://doi.org/10.1093/0199271941.001.0001
  • He, Z., Bian, T., Yang, Z., Li, H., Jiang, N., Cui, J., Jia, B., Wang, Y., Huang, S., & Zhu, Y. (2026). ViT-H: Generative video-to-humanoid control with motion tracking feedback. arXiv. https://arxiv.org/abs/2610.32966
  • Merleau-Ponty, M. (2012). Phenomenology of perception (D. A. Landes, Trans.). Routledge. (Original work published 1945). https://doi.org/10.4324/9780203720714
  • Prpa, M., Fdili Alaoui, S., Schiphorst, T., & Pasquier, P. (2020). Articulating experience: Reflections from experts applying micro-phenomenology to design research in HCI. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3313831.3376664
  • Rodrigues, L. A., & Putze, F. (2026). Local temporal contrasts on human motion data induce a latent geometry. arXiv. https://arxiv.org/abs/2610.21630
  • Rotstein, N., Tewel, Y., Tessler, C., Maron, H., Chechik, G., Litany, O., Kasten, Y., & Atzmon, Y. (2026). Joints, camera, action! Co-generating 3D motion and video for arbitrary rigs. arXiv. https://arxiv.org/abs/2610.28632