Almost every AI system that predicts or generates human movement is built around a single fixed skeleton — a specific list of joints, in a specific order, connected in a specific way. Feed it a body described differently and it fails, which is why a model trained on one motion capture dataset usually cannot be used on another without retraining. A paper accepted to ECCV 2026 called EquiFusion changes this by treating the skeleton as information the model receives rather than a structure baked into it. The result is a system that can work with body descriptions it has never seen, and can keep working when part of the body is hidden from view.

The hidden assumption

When researchers build a motion model, they must decide how a body will be represented. The usual answer is a skeleton: a list of joints — pelvis, spine, left shoulder, left elbow, and so on — with a fixed ordering and a fixed map of which joint connects to which.

That decision then becomes invisible infrastructure. The model's internal machinery is shaped around that exact list. Joint number 14 means "left knee" everywhere inside the system, permanently.

The problem is that different motion capture datasets use different skeletons. One might track 22 joints, another 17, another 52 with detailed hands. They order them differently and define them slightly differently. So a model trained on one is not merely inaccurate on another — it is receiving what amounts to a different language while assuming it is still reading the first one.

The usual workaround is conversion: translate every dataset into one standard skeleton before training. This works, roughly, but loses information, introduces errors, and leaves the underlying rigidity intact. The model still knows exactly one body.

What EquiFusion does differently

EquiFusion's move is to hand the skeleton to the model as an input — an explicit description supplied alongside the movement data, saying which joints these are and how they connect — rather than building the model around one skeleton in advance.

To make this work, the internal machinery has a property called permutation equivariance. In plain terms: the model's computations do not depend on the order in which joints arrive. Shuffle the list and describe the connections accordingly, and the model behaves consistently. It has no fixed idea that joint 14 is the left knee, because it reads which joint is which each time.

This is a change in what the model has learned. Instead of learning "how this particular skeleton moves," it learns something closer to "how connected bodies move" — a pattern it can apply to a body description it has never encountered.

The reported results support this. The model generalises across datasets to skeletons it was never trained on, and does so while being up to 75% smaller than systems built for one skeleton, with faster training. Dropping the assumption made it smaller rather than more complicated, which is usually a sign that the assumption was doing work the model no longer needs to do.

The capability that matters most

The consequence worth attention is what happens with incomplete data.

Because the model reads which joints it is being given, it does not require the full set. Hand it only the joints that are visible and it predicts from those — with no retraining, no special mode. It also works in reverse: ask it to generate movement for particular limbs while holding others fixed.

For anyone working with real bodies in real spaces, this addresses the ordinary condition rather than an edge case. Bodies fold, turn away, go to the floor, and stand behind one another. Cameras lose sight of joints constantly. A model that requires a complete clean skeleton fails in a studio; a model that works from whatever is visible does not.

There is a parallel development in wearable sensing. A July 2026 paper introduced a system reconstructing full-body movement from whatever consumer devices someone happens to be wearing — phone, watch, glasses, insoles, in any combination, robust to whichever are missing. Same underlying idea, applied to sensors instead of joints: stop requiring the world to arrive in a fixed format.

What this does not solve

It is worth being precise about the limits, because "works with any body" invites a larger reading than the result supports.

EquiFusion is agnostic about how a body is described, not about what a body is. It handles different joint lists. It does not thereby understand that a tall long-limbed mover and a compact one solve a movement differently, or that one performer's habitual patterns differ from another's. Those differences show up in the movement data, not the skeleton format.

More fundamentally, this remains prediction of joint positions over time. It says nothing about muscular effort, about whether a gesture is held or released, or about what the movement felt like to perform. Those live in signals a skeleton does not carry — a boundary this series has returned to repeatedly, most directly in why a robot can see your hand but cannot feel it.

What has changed is narrower and still real. One assumption that had quietly limited these systems — that a body must arrive in the format the model expects — turns out to be removable. Systems get more flexible, more compact, and considerably more usable outside the lab. The deeper question of what a skeleton leaves out is untouched, and it is the one this platform keeps returning to in its account of where movement quality is actually constituted.

References

Curreli, C., Hofherr, F., Muhle, D., Saroha, A., Marin, R., & Cremers, D. (2026). EquiFusion: Kinematics-agnostic human motion prediction via equivariant latent diffusion. arXiv:2607.10984. https://arxiv.org/abs/2607.10984

Towards real-world wearable motion reconstruction. (2026). arXiv:2607.09780. https://arxiv.org/abs/2607.09780