A new system lets AI control each body part separately, like a puppeteer with independent strings. It works impressively well — and in doing so, it shows exactly what a living body does that a puppet cannot.
Here is a task that sounds simple and turns out to be surprisingly hard for AI. Generate a movement where a person waves with the right hand, scratches their head with the left, shifts their weight to one hip, and nods — all at slightly different times, each body part on its own schedule.
Until recently, AI movement systems handled instructions at the level of the whole body or whole actions: "walk," "wave," "sit down." Asking for several different things happening in different body parts at different times, all at once, was beyond them. This month, a system called FrankenMotion cracked it — the name a knowing nod to Frankenstein's monster, assembled from parts. It can control the body part by part, each limb following its own detailed instruction on its own timing.
It works remarkably well. And the way it works — and the place where a movement practitioner would immediately sense something missing — tells us something important about the difference between a body and a machine that imitates one.
The Power of Parts
Let's give FrankenMotion its due, because the achievement is real and useful.
Being able to control body parts independently is genuinely valuable. A choreographer describing a complex phrase often does think in terms of parts: "the right arm sweeps up while the gaze follows and the left hand stays anchored." Sign language, conducting, many technical dance vocabularies, and countless everyday coordinations involve different body parts doing genuinely different things at once. An AI that can follow that kind of fine-grained, multi-part instruction is a more expressive and precise tool than one limited to whole-body actions.
To build it, the researchers had to solve a data problem: existing motion datasets didn't describe movement at the part level in enough detail. So they used an AI language model to generate fine-grained, part-by-part, moment-by-moment annotations of movement — a "part vocabulary" the system could learn from. It learned to assemble complex whole-body movement by composing these independently specified parts. Hence Frankenstein: the whole built up from separately controlled pieces.
What the Puppet Cannot Do
Now stand a movement practitioner in front of this system, and here is what they would notice — not as a technical bug, but as a difference in kind.
In a living body moving well, the parts are not independent. When you reach for something across a table, it is not that your arm moves while the rest of you holds still. Your reach travels through your whole body: your spine lengthens, your weight shifts, your opposite side subtly counterbalances, your breath adjusts. The reach is not performed by the arm; it is performed by the whole body, expressed through the arm. Pull on any one part, and the whole system responds, because the parts are connected — not just mechanically, but in the organisation of the movement itself.
This is one of the first things serious movement training teaches, and it goes by names like connectivity, integration, or whole-body organisation. A trained mover's arm gesture carries the whole body in it; a beginner's arm gesture is "just the arm," disconnected, effortful, somehow dead-looking even when it hits the right position. The difference — visible to any trained eye, invisible to a position-tracker — is precisely whether the movement is integrated through the body or assembled from isolated parts.
And that is exactly the line FrankenMotion sits on. It composes movement from independently controlled parts. It is, by design, a Frankenstein: parts stitched into a whole. It can make each part do the right thing at the right time. What it does not do — cannot do, by its very architecture — is generate the whole-body connectivity through which a living body's parts are never really separate in the first place.
Why "Assembled" and "Integrated" Are Different in Kind
It would be easy to think this is just a matter of degree — that a good-enough part-composition system would eventually look integrated. But the difference is deeper than that, and worth being precise about.
When movement is genuinely integrated, the relationship between the parts is not something added on top of the parts; it is where the movement lives. The counterbalance, the sequencing of impulse through the spine, the way weight travels and the breath supports — these are not extra instructions layered onto independently moving limbs. They are the movement's actual organisation, the thing from which the parts' motions follow. Integration is not parts-plus-coordination. It is a different starting point: the whole first, the parts as its expression.
Assembling parts and then coordinating them, however skillfully, starts from the opposite end — parts first, whole as their sum. And a sum of parts, no matter how well timed, does not reconstitute the through-body connectivity of a movement organised as a whole. You can get closer and closer to the right appearance. But the underlying organisation is inverted, and a trained perceiver feels it.
This is not a knock on FrankenMotion, which is excellent at what it sets out to do, and genuinely useful for the many real cases where articulated part-control is what you want. It is an observation about what part-level composition, as an approach, structurally is — and therefore about what it leaves for whole-body somatic organisation to be.
Why This Matters
The reason this distinction is worth drawing carefully is that it points to something a movement practitioner knows that AI research is still discovering: that the whole is not built from the parts, but the other way around.
As AI movement systems become more capable, they will become better and better at the parts — finer control, more precise timing, more articulate composition. This is real progress and it will produce genuinely useful tools. But the integration — the living connectivity through which a body's parts are never truly separate — is not waiting at the end of the part-control road. It is a different thing, organised from the whole down, and it is exactly the dimension that somatic practice cultivates and that trained perception can assess.
There is a quiet lesson here about collaboration between human movement expertise and AI. The machine, for now, assembles. The living body integrates. A partnership that understood the difference — using the machine's compositional precision where articulation is wanted, and the human's whole-body organisation where integration is essential — would be getting the best of both, precisely by not confusing one for the other.
Continue through the archive
For connected context, read This Week in Motion AI: Parts, Wholes, and the Brain–Cerebellum Split and Community Digest: Week of 30 June – 6 July 2026.
References
Li, C., et al. (2026). FrankenMotion: Part-level human motion generation and composition. arXiv:2601.10909. https://arxiv.org/abs/2601.10909
Bainbridge Cohen, B. (1993). Sensing, feeling, and action: The experiential anatomy of Body-Mind Centering. Contact Editions.