Community Scout digest — 6–12 May 2026. Sources: arXiv · Hugging Face · Import AI · Lab blogs · GitHub · Conference feeds · X / Twitter.
1. arXiv
An Elastic Shape Variational Autoencoder for Skeleton Pose Trajectories (May 9) Rahman, A., Kumar, S., Barnes, L. E., & Srivastava, A. arXiv:2605.09231. https://arxiv.org/abs/2605.09231 Manifold-aware VAE for skeletal sequences using elastic shape representation; produces geometrically meaningful latent spaces for movement. Strong contribution to motion manifold learning.
ActCam: Zero-Shot Joint Camera and 3D Motion Control for Video Generation (May 7) El Khalifi, O., et al. arXiv:2605.06667. https://arxiv.org/abs/2605.06667 Simultaneous character motion and camera trajectory control in video generation without task-specific training. Opens compositional motion-conditioned video control for performance documentation.
PersonaGesture: Single-Reference Co-Speech Gesture Personalization for Unseen Speakers (May 7) Zhang, X., et al. arXiv:2605.06064. https://arxiv.org/abs/2605.06064 Personalises gesture generation to a new speaker from one reference clip by separating identity from utterance dynamics. Single-calibration-clip personalisation directly applicable to somatic AI pipelines.
Unison: Harmonizing Motion, Speech, and Sound for Human-Centric Audio-Video Generation (May 9) Cheng, S., Zhang, J., Song, Q., et al. arXiv:2605.08729. https://arxiv.org/abs/2605.08729 Semantic-guided harmonisation across motion, speech, and sound in joint audio-video generation. Relevant to multimodal performance synthesis where body movement, voice, and soundscape must cohere.
OmniHumanoid: Streaming Cross-Embodiment Video Generation with Paired-Free Adaptation (May 12) Song, Y., Deng, X., Yang, P., Wang, Y., & Shou, M. Z. arXiv:2605.12038. https://arxiv.org/abs/2605.12038 Motion transfer across diverse humanoid embodiments using unpaired lightweight adaptation; streaming inference. Generalises motion to non-standard bodies without retraining — relevant to diverse body representation in somatic AI.
2. Hugging Face Daily Papers
Trending papers this week were dominated by LLM reasoning and multimodal vision-language work. No dedicated dance/pose/embodied-AI papers reached the top 10 in the May 6–12 window. Motion-adjacent: HumanNet (arXiv:2605.06747) — 1M-hour human activity video dataset — appeared in the HuggingFace papers feed. https://huggingface.co/papers/2605.06747
3. Import AI #456 (May 11 🚩)
Clark, J. (2026, May 11 🚩). Import AI 456: RSI and economic growth; radical optionality for AI regulation; and a neural computer. https://importai.substack.com/p/import-ai-456-rsi-and-economic-growth
Topics: recursive self-improvement and economic growth modelling; AI regulation frameworks ("radical optionality"); a neural computer architecture (learned unified compute/memory/interface). No motion, embodied AI, dance, or body movement content this issue.
4. Lab Blogs
Meta FAIR — Seamless Interaction (original post July 2025; paper arXiv:2506.22554 continues to circulate widely this week) ai.meta.com/blog/seamless-interaction-natural-conversational-dynamics/ — Family of audiovisual dyadic motion models trained on a 4,000-hour, two-person interaction dataset (4,000+ unique participants). Models generate full-body gestures and active listening behaviours from dual audio streams; reactive to visual cues (smile mirroring, mutual gaze). The Seamless Interaction Dataset is the largest known dyadic human behaviour dataset released to date. Directly relevant to conversational embodied AI and partnered movement research.
5. GitHub Trending
MACE-Dance (updated May 7) — https://github.com/Youlixiya/MACE-Dance — Music-driven dance video generation using Mixture-of-Experts architecture (Motion Expert + Appearance Expert); updated to v2 with improved kinematic plausibility and expressiveness evaluation. 3D motion generation from music, then rendered to video.
No additional dance/motion-capture/somatic-AI repos in Python trending this week.
6. Conference News
HuMoGen @ CVPR 2026 — 1-page abstract deadline passed May 10. Workshop at CVPR Denver, June 3–7. Programme to be announced. https://humogen.github.io/
ICRA 2026 — Vienna, June 1–4. Full programme live. WBCD bimanual manipulation competition begins. https://2026.ieee-icra.org/
ECogS 2026 — Okinawa, OIST, November 9–13. Theme: Embodied Cognition and AI. Call for abstracts open. https://x.com/DrTomFroese/status/2026992275742761202
7. X / Twitter
@AIatMeta — Meta FAIR's Seamless Interaction models demonstrated generating full-body listening and speaking gestures synchronised to two-person audio. Dyadic reactivity (smile mirroring, gaze following) across speech turns. Strong signal for embodied conversational AI. https://x.com/AIatMeta/status/1938641490512851290
@EmbodiedAIRead — Continued coverage of VLA research landscape: 164 ICLR 2026 submissions in the Vision-Language-Action category (vs. 9 the prior year). Embodied action models are scaling at the same rate as language models did in 2022. https://x.com/EmbodiedAIRead/status/1978957826040172579
"somatic AI" and "dance technology" site:x.com searches for May 6–12 returned no additional organic on-topic posts beyond the above. Sponsor/Ad results excluded.