Community Digest: Week of 19–25 May 2026

CVPR 2026 opens in Denver, NVIDIA releases a full humanoid motion stack, and the motion dataset field takes a taxonomic turn

From the arXiv

Skinned Motion Retargeting with Spatially Adaptive Interaction Guidance (arXiv:2605.19355, May 19) KAIST Visual Media Lab introduces geometry-aware retargeting that preserves interaction semantics (self-contact, proximity, object contact) across bodies with significantly different proportions. Uses a Transformer-based anchor refinement strategy with differentiable soft projection to target geometry. Key advance: retargeting that is semantically faithful rather than just positionally close. Directly relevant to identity-conditioned generation for specific practitioner bodies. → https://arxiv.org/abs/2605.19355

RoMo: A Large-Scale, Richly Organised Dataset and Semantic Taxonomy for Human Motion Generation (arXiv:2605.26241, May 25) Researchers from ANU, Roblox, Stanford, and Rutgers release a large-scale motion dataset with a three-level semantic taxonomy and detailed captions per sequence. Taxonomy-aware filtering removes static and low-quality sequences. Per-category evaluation metrics reveal model strengths/weaknesses that global FID scores obscure. Trained models achieve state-of-the-art on standard benchmarks with improved handling of subtle, complex prompts. → https://arxiv.org/abs/2605.26241

PhyGenHOI: Physically-Aware 4D Generation of Dynamic Human-Object Interactions (arXiv:2605.30268, May 28) Ben-Ishu et al. address 4D Human-Object Interaction generation: given a static 3D human and a target object (represented as 3D Gaussian Splats), synthesise a dynamic scene where the human engages the object through actions (punching, kicking, grasping) consistent with physics. The framework models the human as a semantic agent driven by a Motion Diffusion Model and the object as a physical agent simulated via the Material Point Method. First paper to couple generative human motion with explicit physical object simulation at this fidelity. → https://arxiv.org/abs/2605.30268

From the Industry

NVIDIA releases full open-source humanoid motion stack — NVIDIA announced general availability of the Isaac Lab robot learning framework alongside six new workflows for Project GR00T (including GR00T-Mimic for motion and trajectory generation) and a full ecosystem of tools for capturing, generating, retargeting, and simulating humanoid motion data at scale. The Isaac GR00T N1.6 vision-language-action model enables full-body humanoid control. This is the most comprehensive open-source release of embodied AI infrastructure to date and significantly lowers the barrier for researchers working on human-to-robot motion transfer. → https://blogs.nvidia.com/blog/robot-learning-humanoid-development/

From Import AI

Import AI #458 (Jack Clark, May 26, 2026) — An unusually reflective issue: a full-length essay adapted from Clark's Cosmos HAI Lab Lecture at the Oxford Institute for Ethics in AI, on how society should process the success of AI as a technology and what "exploring the future rather than retreating from the present" might look like across individuals and institutions. Paired with a speculative fiction piece imagining a positive singularity scenario. More philosophical than technical — relevant to the question of how AI development should be oriented, including in creative and somatic domains. → https://jack-clark.net/2026/05/26/import-ai-458-reckoning-with-the-future-and-a-singularity-story/

From X / Social

@UmarIqb (Umar Iqbal, NVIDIA) — "NVIDIA just released a whole ecosystem for human(oid) motion and robot learning from human data 🚀🦾 Data, as we all know, is the key to scaling AI models. To accelerate the field of Embodied AI, we have open-sourced a full stack of models and tools to capture, generate, retarget, and simulate human(oid) motion data at scale." The announcement generated significant discussion among embodied AI researchers.

@qineng_wang — reminder that FMEA Workshop challenge test-set submission deadline was May 25. The Foundation Models Meet Embodied Agents challenges at CVPR 2026 cover VLMs, LLM planning, and VLA policies; cash prizes per track. The challenge window closing marks the pre-conference phase.

Conference Notes

CVPR 2026 opens June 1, Denver, Colorado — The 43rd IEEE/CVF Conference on Computer Vision and Pattern Recognition begins its main program June 1. With 4,090 accepted papers (25% acceptance rate, 42% increase in submissions year-on-year), this is the largest CVPR to date. Human motion papers accepted include Superman (arXiv:2602.02401), MotionPRO (arXiv:2504.05046, pressure-based motion capture), and multiple physics-aware generation papers. The HuMoGen Workshop and Foundation Models Meet Embodied Agents Workshop both run as satellite events. This is the first CVPR with a dedicated oral track for human motion generation.

References

Ben-Ishu, O., et al. (2026). PhyGenHOI: Physically-aware 4D generation of dynamic human-object interactions. arXiv:2605.30268. https://arxiv.org/abs/2605.30268

KAIST Visual Media Lab. (2026). Skinned motion retargeting with spatially adaptive interaction guidance. arXiv:2605.19355. https://arxiv.org/abs/2605.19355

[RoMo authors]. (2026). RoMo: A large-scale, richly organised dataset and semantic taxonomy for human motion generation. arXiv:2605.26241. https://arxiv.org/abs/2605.26241