3D Intelligence Report – June 11, 2026
Five verified finds. The day's red thread is hard to miss: Gaussian splats crossed from research into infrastructure this week (Apple), while the research front moved on to predicting the future of a scene, not just reconstructing its present (Envision4D).
Every link below was fetched and verified on June 11, 2026, the day this report went out.
Envision4D: Envisioning Visual Futures via Feed-forward 4D Gaussian Splatting for Autonomous Driving
Qi Song, Yifei He, Chi Zhang, Zheng Fu, Xuhe Zhao, Mengmeng Yang, Kun Jiang, Rui Huang, Diange Yang (Tsinghua University, CUHK-Shenzhen)
6 future frames, zero future poses
Feed-forward 4D Gaussian splatting that extrapolates the future of a driving scene without ground-truth future poses or motion priors. At 6 future frames it matches what STORM delivers at 2, so it resists the error accumulation that kills most extrapolation methods. Reconstruction and prediction in one forward pass: the geometry community's answer to world models.
Driving teams keep treating reconstruction and prediction as two separate systems. Envision4D folds them into one forward pass: it builds the 4D scene and extrapolates what happens next, with no future poses and no motion priors. The part I would steal is the progressive training that stops unsupervised motion learning from compounding its own errors. If you work with mobile mapping data, pose-free is the phrase to watch this year.
Research Engineer / Scientist (3D Reconstruction)
$250K-$350K + equity
World Labs published a taxonomy of world models eight days ago, and this posting reads like its staffing plan: renderers need geometry people. They want multi-view reconstruction, pose estimation, robust losses, differentiable rendering. My read: classic photogrammetry skill got rebranded as frontier AI, and the salary moved with the label. Lead your application with a pipeline that survived uncontrolled capture, not a benchmark score.
Four funding doors, four regions
next deadline in 24 days
-
Asia deadline 2026-07-05 · 24 days leftACCV 2026 call for papers (Osaka, Japan)
Springer LNCS proceedings, conference Dec 14-18, 2026
The post-CVPR window is exactly for this. The 3D work that barely missed the CVPR cut deserves a second, sharper version, not a drawer. Registration closes July 3, papers July 5, and 3D computer vision is named in the topics list.
-
Americas deadline 2026-07-13 · 32 days leftNSF X-Labs, Topic 2: Scientific Instrumentation for Sensing and Imaging
part of a $1.5B initiative, large multiyear milestone-based awards
A $1.5B bet on independent teams building instruments, not writing papers. Lidar, computational imaging and 3D capture people, this one is aimed at you. The milestone-based format favors builders, and written proposals close July 13.
-
France deadline 2026-09-03 · 84 days leftFrance 2030 – PEPR Climat TRACCS, appel a projets 2026
EUR 500K-1M per project (+ EUR 300-400K early-career fellowships)
Climate modeling sounds far from 3D until you read axis 1: synergies between observations and modelling. That is Earth observation pipelines, and our community already runs them. Phase 1 is a two-page letter of intent, the cheapest ticket in French research funding right now.
-
Europe deadline 2026-09-10 · 91 days leftEurostars Call 11 (Eureka)
national co-funding for SME-led R&D consortia (rates vary by country)
The rare program where the small company leads and the consortium forms around it. Submissions open July 9 and close September 10 at 14:00 CET. September sounds far away, but the partner search is the slow part. Start the conversations now.
PyTorch Geometric v2.8.0
v2.8.0
torch-cluster and torch-spline-conv consolidated into pyg-lib 0.7 (the knn/radius graph ops every PointNet++-style 3D pipeline depends on), wheels for PyTorch 2.9-2.12 and Python 3.10-3.14, LLM+graph workflows (synthetic QA generation, RelBench graph construction from relational databases), faster to_dense_batch.
Every PointNet++-style pipeline leans on torch-cluster for knn and radius graph ops, and that dependency has broken more student installs than any other in my courses. 2.8 folds it into pyg-lib: one less wheel to chase on Windows. If you teach or ship 3D deep learning, this is the version bump to schedule, not skip.
RealityKit ships native 3D Gaussian splat rendering at WWDC26; Apple Maps Flyover moves to splats
GaussianSplatResource and GaussianSplatComponent give splats first-class API status in RealityKit (position, scale, rotation, opacity and spherical harmonics buffers, format-agnostic), demoed on Vision Pro with sample code. Apple Maps Flyover is moving to 3DGS. When the largest consumer platform makes a representation an OS primitive, the format debate is settled.
When Apple gives a representation first-class API status, the format war is over. Splats now render natively in RealityKit, buffers in, photoreal scene out, and Apple Maps Flyover is moving to the same representation. If you capture reality for a living, your output just became an operating system primitive. The question is no longer whether splats are production-ready, it is who fills the pipeline.
Get the next report in your inbox
Five verified finds, my take on each, one short email a day.
Five verified finds with my take, one short email a day.
