Three releases from 1–7 September 2026 broaden what motion can teach a generative system. Kirin reconstructs and generates quadruped motion from in-the-wild video; MudraGen uses geometric supervision to generate interacting two-hand mudras from Bharatanatyam; and Object Concepts Emerge from Motion uses motion boundaries in raw video to teach a static-image encoder about individual objects. The shared advance is representational: motion is becoming a source of structure for species, culturally specific gesture and object identity—not merely an animation output.

Kirin is listed for ECCV 2026 and MudraGen is accepted for an ACM Journal on Computing and Cultural Heritage special issue. The object-learning paper is a preprint. Reported results remain author-reported.

Kirin builds animal motion priors from videos that already exist

Paper: Zhao, B. N., Pan, Z., Rehg, J. M., Wu, J., & Wu, S. (2026, 1 September). Kirin: Animal motion generation from in-the-wild video. arXiv:2609.01823. ECCV 2026.

Animal motion datasets are difficult to collect with laboratory motion capture across many species. Kirin instead reconstructs 3D sequences from large collections of ordinary animal video and pairs video, text and motion in the AiM3D dataset. A generation model then conditions on both an image and text, while an image-to-3D system rigs and animates generated assets.

Why it matters. The project tests whether a motion prior can scale without assuming a human skeleton or studio capture. Reconstruction errors, species imbalance and the behaviour selected by online video remain important limitations. The archive’s analysis of body-agnostic motion systems explains why removing one fixed body assumption does not remove dataset bias.

MudraGen treats hand geometry as cultural structure

Paper: Kamble, J. K., Mukhopadhyay, J., Roy, D., & Das, P. P. (2026, 3 September). MudraGen: Geometrically supervised generation of interacting two-hand mudras for preserving Indian classical dance heritage. arXiv:2609.03415.

MudraGen generates images of Samyukta Hasta Mudras, coordinated two-hand gestures used in Bharatanatyam. Because canonical Sanskrit definitions do not necessarily provide the detailed textual descriptions expected by text-to-image systems, the method adds three geometric objectives: 3D keypoint alignment, offsets between the hands and shape consistency. The authors report improvements in visual realism, anatomical correctness and hand-pose structure.

Why it matters. The paper recognises that a culturally specific gesture cannot be recovered from generic prompt language alone. Geometry can protect visible coordination, but cultural meaning also depends on pedagogy, context and practitioner authority. A preservation tool should therefore support transmission without claiming that a generated image contains the practice itself. This extends the Lab’s account of the cultural vocabulary gap in motion AI with a concrete low-resource case.

Motion boundaries teach a still-image model about object identity

Paper: Li, B., Wang, X., Wu, X., Li, Z., Yang, Y., & Wang, N. (2026, 3 September). Object concepts emerge from motion. arXiv:2609.04348.

The method clusters optical flow to create pseudo-instance masks from raw video, then uses pixel-level metric learning to train a single-image encoder. It begins with 195 million pseudo-labelled frames drawn from 7,163 hours of driving and web video and expands to 421 million frames through motion-verified self-training. The authors report competitive or stronger transfer on depth, 3D detection, occupancy prediction and planning tasks.

Why it matters. Movement supplies a grouping cue: pixels that move together often belong to one object. That cue can teach a model about individuality without hand-labelled masks. It is powerful but not universal—stationary objects, articulated bodies and camera motion complicate the rule. Motion offers evidence for objecthood, not a complete definition.

The combined signal

This week shows motion working at three scales: as a species-specific prior, a culturally constrained hand relation and a visual cue for object identity. The opportunity is to build representations from meaningful relations rather than generic frames. The obligation is to preserve provenance: reconstructed animal motion, generated mudras and flow-derived object masks are all inferences, not direct recordings of behaviour, heritage or identity.

References

Kamble, J. K., Mukhopadhyay, J., Roy, D., & Das, P. P. (2026). MudraGen: Geometrically supervised generation of interacting two-hand mudras for preserving Indian classical dance heritage. arXiv:2609.03415. https://arxiv.org/abs/2609.03415

Li, B., Wang, X., Wu, X., Li, Z., Yang, Y., & Wang, N. (2026). Object concepts emerge from motion. arXiv:2609.04348. https://arxiv.org/abs/2609.04348

Zhao, B. N., Pan, Z., Rehg, J. M., Wu, J., & Wu, S. (2026). Kirin: Animal motion generation from in-the-wild video. arXiv:2609.01823. https://arxiv.org/abs/2609.01823