Community Digest: Week of 2–8 June 2026

CVPR 2026 concludes with awards, the HuMoGen workshop convenes its third edition, and contact-aware mocap goes markerless

From CVPR 2026 (Denver, June 3–7)

Best Paper Award: D4RT — The CVPR 2026 Best Paper went to "Efficiently Reconstructing Dynamic Scenes One D4RT at a Time" (Zhang et al., Google DeepMind / UCL / Oxford). D4RT reconstructs the geometry and motion of dynamic 4D scenes from video using a single unified transformer — replacing the traditional multi-model pipeline (separate depth, flow, and camera-pose models) with one latent representation answering all queries. New state of the art across every 4D reconstruction benchmark. → https://d4rt-paper.github.io/

Award Candidate: MAMMA (Oral, top ~1.75%) — "Markerless Accurate Multi-person Motion Acquisition" (Velasquez et al., Max Planck Institute for Intelligent Systems / CMU) achieves marker-grade motion capture accuracy from multi-view video without markers — including for closely interacting multiple bodies. The transformer-based dense landmark estimator predicts contact probabilities alongside landmarks, making physical contact between bodies a modelled signal rather than a failure mode. Code public. → https://mamma.is.tue.mpg.de/

HuMoGen Workshop (3rd edition) — The Human Motion Generation workshop convened with a focus on simulation, animation, and VR applications, plus interaction motion synthesis (human-scene, human-human, human-object). Notable presentation: Sebastian Starke's AI4AnimationPy, an open-source Python framework for character animation research — lowering the entry barrier for researchers outside the games industry. → https://humogen.github.io/

PhysHuman Workshop (1st edition) — The inaugural physics-based human modelling workshop brought together computer vision, biomechanics, and simulation researchers around the question of how bodies should move under real-world physical constraints. The workshop's existence confirms physics-grounded human modelling as an established research community within mainstream CV. → https://physhuman.github.io/

From the arXiv

KV-Control: Parameter-Efficient K/V Injection for Trajectory-Controlled Text-to-Motion (arXiv:2606.05624, June 4) Sun et al. inject trajectory constraints directly into the key/value streams of a frozen text-to-motion transformer's self-attention — no retraining, no output-side enforcement. The character follows a sketched path or hits an end-effector target while preserving the gait and style described in language. The frozen-backbone + compact-injection pattern continues to consolidate as the dominant control paradigm. → https://arxiv.org/abs/2606.05624

From Import AI

Import AI #460 (Jack Clark, June 8, 2026) — A shorter issue than usual; Clark notes he spent the week attending the 2026 Bilderberg conference. The condensed edition continues threads from #459 on AI oversight infrastructure and the economics of the AI buildout. → https://importai.substack.com/

From X / Social

@gh_marjan (Marjan Ghazvininejad, Meta) — Meta hosted its CVPR 2026 networking mixer in Denver (June 5), one of the larger industry-community events of the conference week. Meta's FAIR motion and embodied AI teams were recruiting heavily — an industry signal that motion AI hiring remains strong.

@abursuc (Andrei Bursuc) — CVPR@Paris ran June 1 as a satellite event for the European community unable to travel to Denver, with poster sessions and plenary talks. The satellite model (regional mirror events around the main conference) is becoming standard for the largest ML conferences.

CVPR community discussion — The conference's 16,000+ submissions and 4,090 accepted papers generated the recurring debate about review quality at scale. Multiple prominent researchers proposed that motion, embodiment, and robotics may warrant their own dedicated conference within 2–3 years if growth continues.

Conference Notes

Next deadlines on the radar:

  • NeurIPS 2026 — abstract deadline passed (May); paper notifications September
  • ECogS 2026 (Embodied Cognitive Science, OIST, Nov 9–13) — abstract submissions expected to open summer; theme "Embodied cognition and AI"
  • MOCO 2026 (International Conference on Movement and Computing) — the practice-oriented movement computing community's annual meeting; check movementcomputing.org for dates

References

Sun, T., et al. (2026). KV-Control: Parameter-efficient K/V injection for trajectory-controlled text-to-motion. arXiv:2606.05624. https://arxiv.org/abs/2606.05624

Velasquez, H. C., et al. (2026). MAMMA: Markerless accurate multi-person motion acquisition. In Proceedings of CVPR 2026. https://mamma.is.tue.mpg.de/

Zhang, C., et al. (2026). Efficiently reconstructing dynamic scenes one D4RT at a time. In Proceedings of CVPR 2026. https://d4rt-paper.github.io/