Two papers published in July 2026 attack the same hidden constraint from opposite ends: motion AI systems are usually locked to one fixed skeleton and one fixed sensor rig. EquiFusion (arXiv:2607.10984, 13 July, accepted to ECCV 2026) makes human motion prediction kinematics-agnostic, treating skeletal structure as an input parameter rather than a baked-in assumption. WHIP (arXiv:2607.09780, 8 July) makes motion reconstruction sensor-agnostic, rebuilding full-body movement from whatever consumer devices a person happens to be wearing. For movement practice, both matter for one reason: they loosen the requirement that a body must match the system's expectations before the system can work with it.

This digest covers three papers from July 2026. Publication dates and venues are given for each; all three are preprints or conference papers with public arXiv records.

EquiFusion: prediction that does not assume a fixed skeleton

Paper: Curreli, C., Hofherr, F., Muhle, D., Saroha, A., Marin, R., & Cremers, D. (2026, 13 July). EquiFusion: Kinematics-agnostic human motion prediction via equivariant latent diffusion. arXiv:2607.10984. Accepted to ECCV 2026.

Nearly every human motion model is trained on one skeletal format — a specific set of joints in a specific order with a specific connectivity graph. Change the skeleton and the model breaks, which is why models trained on one motion capture dataset typically cannot be applied to another without retraining.

EquiFusion removes this constraint by making the skeleton's connectivity an explicit input to the model rather than an implicit structural assumption. Its permutation-equivariant architecture means internal computations do not depend on joint ordering or graph structure. The authors report state-of-the-art results on major benchmarks while being up to 75% more compact than kinematics-specific methods, with faster training and inference.

The capabilities this unlocks are the interesting part. Because the model is agnostic to which joints it receives, it supports zero-shot prediction from partial or occluded observations — predicting movement when some of the body is not visible — and targeted limb generation, generating movement for specified body parts. Neither requires retraining.

Why this matters for somatic AI: Occlusion and partial observation are the normal condition of real practice spaces, not an edge case. A model that degrades gracefully when it cannot see the whole body is far more usable in a studio than one that requires complete, clean joint data. The kinematics-agnostic property also points toward body-specific work: if skeletal structure is a parameter, a practitioner's own proportions become something the model accepts rather than something it must be retrained around. That said, this is prediction of joint trajectories — it does not touch the felt or muscular dimension this platform tracks in its analysis of where movement quality is physically constituted.

WHIP: reconstruction from whatever you happen to be wearing

Paper: Towards real-world wearable motion reconstruction. (2026, 8 July). arXiv:2607.09780.

Most wearable motion capture research assumes a fixed sensor configuration — a full IMU suit, or a headset-plus-controllers rig. These assumptions do not survive contact with everyday conditions, where a person might have a phone in a pocket, a watch on one wrist, and nothing else.

The authors argue motion capture should prioritise unobtrusive consumer devices — smartphones, smartwatches, smart glasses, smart insoles — and study how they combine. They contribute a large-scale multimodal dataset synchronising consumer-grade sensors with ground-truth 3D motion across 50 activities spanning everyday tasks, sports and social interaction, and WHIP, a generative baseline model that reconstructs motion from arbitrary subsets of available sensors, handling missing modalities and producing physically plausible output.

Why this matters for somatic AI: Cost and intrusiveness are the two barriers that have kept rigorous movement capture inside well-funded labs, a point this platform examined when markerless multi-person capture reached marker-grade accuracy. A system that works from a watch and a phone lowers the barrier further, and the "arbitrary subset" design means a practitioner is not locked out by owning the wrong combination of devices. The limitation is unchanged: these are inertial and positional signals describing where the body is, not signals describing muscular effort or intention.

Two2Four: transferring human movement onto a four-legged body

Paper: Zargarbashi, F., Qiu, Z., Agrawal, D., Coros, S., Sumner, R. W., Guay, M., & Buhmann, J. (2026, 28 July). Two2Four: Generative quadruped puppeteering from human motion. arXiv:2607.26108. DisneyResearch|Studios Switzerland and ETH Zürich.

Two2Four maps ordinary human motion onto quadruped animals. A two-stage diffusion model trained purely on quadruped motion data, combined with a structured conditioning and inpainting strategy, supports walking, running, jumping, sitting and lying, plus fine-grained control such as head movement and individual limb puppeteering. The authors report improved realism and controllability over existing retargeting approaches, targeting animation and virtual production.

Why this matters for somatic AI: This is cross-morphology transfer at its most extreme — two limbs to four, an entirely different body plan. It sharpens a question this platform has followed since examining what is preserved and lost when movement is transferred between bodies: what exactly travels across such a transfer? Not the biomechanics, which are incompatible. What transfers is something more like intention or expressive shape — the performer's intent interpreted through a body that could never execute it literally. That is a legitimate and interesting form of translation, and it is explicitly not a claim that the animal movement carries the human's felt experience.

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

Zargarbashi, F., Qiu, Z., Agrawal, D., Coros, S., Sumner, R. W., Guay, M., & Buhmann, J. (2026). Two2Four: Generative quadruped puppeteering from human motion. arXiv:2607.26108. https://arxiv.org/abs/2607.26108