Article written and manifest updated. Here is the full digest for the record:
Community Scout Digest — 15–21 April 2026
1. arXiv
Re²MoGen: Open-Vocabulary Motion Generation via LLM Reasoning and Physics-Aware Refinement Zheng et al. (2026-04-20). arXiv:2604.17807. https://arxiv.org/abs/2604.17807
Three-stage pipeline: Monte Carlo tree search for LLM-driven keyframe planning → spatiotemporal pose completion → physics-aware RL refinement. Generates semantically consistent, physically plausible character motion from open-vocabulary text descriptions beyond the training distribution. Directly relevant to choreographic ideation and somatic movement specification by language.
HumanScore: Benchmarking Human Motions in Generated Videos Fang, Xiang, Tan, Schuetz, Delp, Fei-Fei & Adeli / Stanford. (2026-04-21). arXiv:2604.20157. https://arxiv.org/abs/2604.20157
Six-metric framework (kinematic plausibility, temporal stability, biomechanical consistency) testing 13 state-of-the-art models. Key finding: visual plausibility and biomechanical accuracy diverge systematically. Provides vocabulary and measurement tools for evaluating whether generated body movement is genuinely credible from a movement-science standpoint, not merely convincing at a glance.
ReImagine: Rethinking Controllable High-Quality Human Video Generation via Image-First Synthesis Sun, Zheng, Li et al. (2026-04-21). arXiv:2604.19720. https://arxiv.org/abs/2604.19720
Learns human appearance quality through image generation first, then uses that prior for temporally consistent video synthesis with SMPL-X pose and viewpoint conditioning. Decouples appearance fidelity from motion coherence — architecturally relevant to body-to-visual pipelines that need high-quality output without full end-to-end video training.
UniCon3R: Contact-aware 3D Human-Scene Reconstruction from Monocular Video Sur et al. (2026-04-21). arXiv:2604.19923. https://arxiv.org/abs/2604.19923
Uses physical contact with environment as a corrective cue in feed-forward monocular body reconstruction, eliminating floating-body artefacts. Real-time spatially aligned 3D body recovery from a single camera — relevant to somatic practice documentation without studio-grade capture infrastructure.
InHabit: Leveraging Image Foundation Models for Scalable 3D Human Placement Kister et al. (2026-04-21). arXiv:2604.19673. https://arxiv.org/abs/2604.19673
Automatic pipeline for generating large-scale synthetic SMPL-X body-in-environment datasets using image foundation models as supervision. Addresses the situated-movement data bottleneck — relevant to building training sets for embodied AI in site-specific or studio performance contexts.
Seedance 2.0: Advancing Video Generation for World Complexity (note: not a dance synthesis model despite the name) Team Seedance. (2026-04-15). arXiv:2604.14148. https://arxiv.org/abs/2604.14148
General multi-modal video+audio generation (text/image/audio/video in → synchronised audio-video out, 4–15 s, 480p–720p). Relevant as background for audio-visual synchrony pipelines; the name is coincidental to the field.
2. Hugging Face Daily Papers
OneHOI: Unifying Human-Object Interaction Generation and Editing (HF Apr 16) https://huggingface.co/papers/2604.14062 — Unified body-pose + grasp + spatial-relationship generation and editing in one model. Relevant to prop-based somatic practice and practitioner co-creation tool UX.
Habitat-GS: A High-Fidelity Navigation Simulator with Dynamic Gaussian Splatting (HF Apr 15) https://huggingface.co/papers/2604.12626 — Embodied AI navigation simulator using dynamic Gaussian splatting for photorealistic spatial training environments. Infrastructure-level relevance; 88 HF upvotes.
3. Import AI #454 (Apr 20)
Jack Clark. https://importai.substack.com/p/import-ai-454-automating-alignment — Covers automated alignment research, HiFloat4, and Kimi K2.5 safety evaluation. No items relevant to motion, dance, embodied physical AI, or body movement this week.
4. Lab Blogs
Gemini Robotics-ER 1.6 (DeepMind, Apr 14) was covered in last week's digest. No new posts on motion, video generation, or embodied AI from DeepMind or Meta AI in the Apr 15–21 window.
5. GitHub Trending
NVIDIA/Isaac-GR00T N1.7 — https://github.com/NVIDIA/Isaac-GR00T — Foundation model for generalist humanoid robot motion (6,835 ★ total, 146 this week). Open-weight whole-body motion intelligence reference. No dance/motion-capture/creative-coding repos in the Python trending list this week.
6. Conference News
MOCO 2026 — ACM proceedings now live; conference opens Thursday 23 April, Cité des Arts Montpellier (23–25 Apr). Programme: six paper sessions, keynotes on bodies in concert motion and adaptive musical entrainment, practice works, doctoral consortium. Primary community gathering for movement × computing research. https://moco26.movementcomputing.org/
NIME 2026 — 23–26 June, Loughborough University London. No new announcements in the Apr 15–21 window.
7. X / Twitter
@qineng_wang (~2026-04-21 🚩) — FMEA Workshop @ CVPR 2026: four embodied AI challenges open for submissions, 300/$200 prizes per challenge. https://x.com/qineng_wang/status/2046655964367958197
"Motion generation" and "dance technology AI" site:x.com searches returned no organic, on-topic posts within the Apr 15–21 window. Sponsored and off-topic results excluded.