Community Digest: Week of 16–22 June 2026

Generalisation by borrowing from video, the evaluation problem comes into focus, and the muscle-level momentum continues

From the arXiv

The Quest for Generalisable Motion Generation: Data, Model, and Evaluation (arXiv:2510.26794, ICLR 2026) Lin, Wang, Lu et al. (S-Lab NTU / SenseTime) present ViMoGen — transferring knowledge from video generation into motion generation to overcome the generalisation bottleneck. Releases ViMoGen-228K (228K motion samples blending MoCap, web-video motions, and video-model-synthesised motion), a dual-branch flow-matching diffusion transformer, and the MBench generalisation benchmark. Strong results on martial arts, dynamic sports, and multi-step prompts. Code + dataset public (MotrixLab/ViMoGen). The video-to-motion knowledge transfer strategy is the field's leading answer to the data-scarcity problem. → https://arxiv.org/abs/2510.26794

PP-Motion: Physical-Perceptual Fidelity Evaluation for Human Motion Generation (arXiv:2508.08179, ACM MM 2025) Zhao et al. (Tsinghua) propose a motion fidelity metric combining physical-feasibility (minimum modification to satisfy physical law, as continuous ground truth) with a human-perceptual fidelity loss — closing the long-standing gap between "physically valid" and "looks real." Aligns better with human judgement than prior metrics. The evaluation problem — how to reliably measure whether generated movement is good — is becoming a central research concern as generation matures. → https://arxiv.org/abs/2508.08179

Ongoing — the muscle-level cluster: Last week's MuscleMimic (arXiv:2603.25544) continues to generate follow-up discussion and forks; the GitHub repo (amathislab/musclemimic) has seen rapid community uptake. The muscle-actuated approach is consolidating as a recognised sub-direction within embodied AI.

From Import AI

Import AI #462 (Jack Clark, June 22, 2026) — "Superpersuasion; self-sustaining AI; paths to ASI." A more speculative, philosophically inflected issue, including the framing question "How religious are beliefs in the singularity?" Clark examines the epistemics of strong AI-future beliefs, the concept of superpersuasion (AI systems that could shift human beliefs at scale), and competing technical paths toward advanced AI. Less directly technical than usual; relevant to the broader question of how AI's trajectory is narrated and believed. → https://importai.substack.com/

From X / Social

@liuziwei7 (Ziwei Liu, NTU) — the S-Lab group behind ViMoGen continued promoting the generalisable-motion line of work, situating ViMoGen within a multi-year programme on unified, generalist motion foundation models. The "borrow generalisation from video generators" strategy drew discussion about whether video-derived motion preserves enough physical and biomechanical validity — a concern directly relevant to somatic applications, where appearance-derived motion may lack felt-quality grounding.

Motion AI community — ongoing discussion following CVPR 2026 about whether human motion / embodiment warrants a dedicated conference, with the evaluation problem (PP-Motion and related work) frequently cited as a sign the subfield has matured enough to need its own venues and standards.

Conference & Community Notes

Two challenges defining the period: The field has clearly moved past "can we generate plausible movement" to two harder questions — generalisation (ViMoGen) and evaluation (PP-Motion). Both are signs of a maturing field. For somatic-AI work, both also surface the same gap: they advance the account of movement-as-observed while leaving movement-as-lived unaddressed.

Venues on the radar:

  • MOCO 2026 (Movement and Computing) — the most practice-aligned venue for somatic-AI work; confirm 2026 dates at movementcomputing.org (carried action item).
  • ECogS 2026 (Embodied Cognitive Science, OIST, Nov 9–13) — theme "Embodied cognition and AI"; abstract window expected summer.
  • NeurIPS 2026 — paper notifications September.

References

Lin, J., Wang, R., Lu, J., et al. (2026). The quest for generalisable motion generation: Data, model, and evaluation. arXiv:2510.26794. https://arxiv.org/abs/2510.26794

Zhao, S., et al. (2025). PP-Motion: Physical-perceptual fidelity evaluation for human motion generation. arXiv:2508.08179. https://arxiv.org/abs/2508.08179