Why Putting Your Dance on Someone Else's Body Is So Hard — and What AI Is Learning About It
Motion retargeting sounds technical. What it's really about is the deeply personal relationship between a body and the movement it knows how to make.
Watch a professional dancer perform a phrase and then ask an amateur to copy it. What you'll notice is that the amateur doesn't produce a scaled-down or imprecise version of the same movement. They produce a different movement — one that is shaped by their body's habits, proportions, strengths, and limitations in ways that have nothing to do with not trying hard enough.
This is not a failure of will or attention. It is a fact about how movement works. Movement is not instructions that any body can execute given sufficient effort. Movement is a relationship between a set of physical intentions and the specific body through which those intentions are expressed. Change the body, and you change the movement — even if the intentions remain the same.
This is the problem that AI researchers call motion retargeting, and it is harder than it sounds.
What Retargeting Is and Why It Matters
Motion retargeting, at its simplest, means taking a movement sequence recorded on one body and transferring it to a different body. It is a core operation in film animation (where a human actor's movement drives a digital character), video games (where a single motion capture session is used to animate many different character types), and increasingly in AI-generated movement (where movements generated for an "average" body need to work on a specific individual's proportions).
The naive approach is geometric mapping: identify the corresponding joints, scale the positions to fit the target body, done. This works reasonably well for small differences in body shape. It breaks down when the bodies are significantly different — when a phrase choreographed for a very tall, long-limbed mover needs to translate to a shorter, more compact body, or when a movement involves self-contact (a hand touching the torso, arms crossed in front of the chest) that would create interpenetration or fail to make contact on a differently proportioned body.
The problem goes deeper than geometry. A reach that is effortful and dramatic on a body with short arms is casual and unremarkable on a body with long arms. A turn that uses the whole body's momentum on a heavy frame may be light and quick on a lighter one. The meaning of the movement — what it communicates, what it costs, what it expresses — is not stored in the joint coordinates. It is produced by the relationship between those coordinates and the specific body that generates them.
What Geometry-Aware Systems Get Right (and Still Miss)
The most recent advance in motion retargeting, from KAIST's Visual Media Lab (arXiv:2605.19355, May 2026), addresses this problem through what researchers call spatially adaptive interaction guidance. Instead of mapping joint positions directly, the system identifies interaction anchors — the points on the body that are functionally significant: where the body touches itself, where it approaches objects or other bodies, where different body parts enter each other's spatial proximity.
These anchors are then adapted to the target body's specific geometry using a Transformer model, ensuring that the functional relationships — the self-contacts, the near-proximities, the spatial intentions — are preserved across the transfer. If the original movement involved a hand resting lightly on the opposite shoulder, the retargeted movement will still involve a hand resting lightly on the target body's shoulder, even if that shoulder is shaped differently and located at a slightly different relative position.
This is a genuine advance. By preserving interaction semantics rather than just position sequences, the system keeps something that matters about the movement intact.
But a somatic practitioner looking at this approach would notice what it still leaves out. The interaction anchors capture what the body is doing to itself and its environment. They do not capture how it feels to do that. The weight of the hand on the shoulder, the soft or firm quality of that contact, the relationship of that contact to the body's overall organisation in the moment — these are the dimensions that make a gesture meaningful in somatic terms, and they are not encoded in spatial proximity maps.
The Body That Is Not a Container
There is a philosophical assumption buried in the geometry problem: that a body is a container that can be swapped out while the movement remains the same inside it. Different container, same movement — just rescale and re-project.
Somatic practice starts from the opposite assumption: the body is not a container for movement. The body is the movement. What a mover brings — their history, their habits, their injuries and compensations, their accumulated physical vocabulary — is not separable from the phrases they produce. A dancer's movement does not exist independently of the body that makes it; it is an expression of that body's relationship with itself and the world.
This means that "transferring" a movement from one body to another is not a well-defined operation in the way that transferring data between hard drives is. What you transfer is not the movement but an approximation — a movement that preserves some of the original's characteristics and necessarily loses others. The question is which characteristics matter most for what purpose.
For film animation, visual plausibility matters most — you want the retargeted movement to look right on screen. For a somatic practitioner using AI tools, felt authenticity may matter most — you want the retargeted movement to feel right in the body, to be biomechanically compatible with the practitioner's own physical organisation. Those are different targets, and current retargeting systems are optimised for the first, not the second.
Why This Is One of the Central Problems for Somatic AI
The retargeting problem is not an edge case in somatic AI research — it is close to the core. Any AI motion generation system that is not trained specifically on the practitioner's own body is, implicitly, doing retargeting: taking movement knowledge learned from a population of bodies and adapting it to this specific body.
The work described above, and the broader research trajectory it represents, is developing the technical tools to do that adaptation more faithfully. Identity-conditioned generation models (like IAM, discussed in the May 1 frontier report) approach the same problem from the generation side — building models that generate movement which is already shaped by a specific body's morphology, rather than generating for a generic body and then retargeting.
Both directions are necessary. Together, they represent the field's gradual recognition that movement is not body-independent data — that the body is not a substrate for a movement that could equally well be performed by any other body of sufficient capability, but an expressive instrument whose particular qualities are inseparable from what it produces.
That recognition is old news in somatic practice. It is new news in AI.
Further Reading
KAIST Visual Media Lab. (2026). Skinned motion retargeting with spatially adaptive interaction guidance. arXiv:2605.19355. https://arxiv.org/abs/2605.19355
Jia, W., et al. (2026). IAM: Identity-aware human motion and shape joint generation. arXiv:2604.25164. https://arxiv.org/abs/2604.25164