3D Intelligence Report – June 15, 2026
**Theme of the day: reconstruction is turning from a per-scene craft into shippable infrastructure.** A splat you predict in one pass instead of cooking per scene. A production reconstruction pipeline that just went Apache 2.0. A world-model lab hiring classical SLAM, not generative researchers. Funding calls pointed straight at embodied, physical AI. The geometry layer is professionalizing.
Every link below was fetched and verified on June 15, 2026, the day this report went out.
Wild3R: Feed-Forward 3D Gaussian Splatting from Unconstrained Sparse Photo Collection
Yuto Furutani, Takashi Otonari, Kaede Shiohara, Toshihiko Yamasaki (The University of Tokyo)
1 forward pass, no per-scene training
This is the photo-tourism problem solved feed-forward: messy public photo collections, mixed lighting, tourists walking through the frame, rebuilt into a Gaussian splat in a single network pass instead of hours of per-scene optimization. It sits where two of the hottest splat threads meet, feed-forward reconstruction and appearance-in-the-wild, and it ships a new WildCity benchmark (200 scenes, 170 lighting conditions, 337,500 images) with released weights.
The interesting move is not the quality, it's the deletion of the optimization loop. Most of us still treat a splat as something you cook per scene; this treats it as something a trained model predicts. Try it on your own awkward captures, the handheld set with people in it, and watch where the appearance handling holds and where it smears. That failure map tells you more than the benchmark does.
Research Engineer / Scientist (SLAM)
$250K-$350K base + equity (disclosed)
Look at what a world-model lab is hiring for: visual-inertial and lidar SLAM, bundle adjustment, factor graphs, Kalman filtering, fused with learned perception. Not a generative researcher, a classical geometry person who can ground a neural world in metric reality. That is the tell. The next phase of spatial AI is persistent, metrically correct mapping, and the people who own that crossover get paid top of band for it. Lead your application with a pipeline that survived real sensor noise, not a clean benchmark.
Four funding doors, four regions
next deadline in 15 days
-
Asia deadline 2026-06-30 · 15 days leftIEEE InGARSS 2026 call for papers (Hyderabad, Dec 1-4)
Conference CFP, IEEE Xplore proceedings + presentation (theme: Digital Earth)
This is the tight one, fifteen days, so it is for work that is already mostly written. The Machine Learning for Digital Earth track and the SAR and remote sensing tracks are squarely on topic, and a December slot in Hyderabad is a real venue to plant a flag in the GRSS community. If you have a remote-sensing or spatial-AI result sitting in a drafts folder, this is the deadline that forces it out.
-
France deadline 2026-09-09 · 86 days leftFrance 2030 i-Demo (appel generique n5, projets structurants de RD&I)
Grant + repayable advance, ~50% (up to 60% small firms), project floor >EUR 2M single / >EUR 4M collaborative
Sector-agnostic deep-tech money, which is exactly why a spatial-AI or digital-twin demonstrator fits the brief. The honest constraint is the project floor: this is for a structured R&D effort moving toward market, not a side experiment. If you have a serious build at TRL 4 to 7, the September window gives you the summer to assemble the budget annexes, which is the part that always eats the time.
-
Americas deadline 2026-09-10 · 87 days leftNSF CISE Future Computing Research (Future CoRe), NSF 25-543
Up to $1M over up to 4 years
The scope covers Robust Intelligence (perception in messy real contexts) and cyber-physical systems, so vision and 3D understanding research has a clean home here. The useful detail: this is a target date, not a hard cliff, the call takes proposals continuously and reviews against September 10. That means the queue is the strategy, get in early and get read early.
-
Europe deadline 2026-10-28 · 135 days leftEIC Pathfinder Challenges 2026 (Horizon Europe)
~EUR 96M indicative, grants up to ~EUR 3-4M per project, TRL 1-4
One 2026 challenge is literally embodied intelligence for AI-powered robotics, which is the academic doorway for the same physical-AI shift the jobs market is signaling. Pathfinder funds early, high-risk science, so this is the call for the idea that is too unproven for the Accelerator. Consortia take months to form; start the partner conversations now, not in October.
NVIDIA 3D Object Reconstruction v0.2.0 (Apache 2.0)
Now Apache 2.0
NVIDIA's stereo-to-mesh pipeline went open and commercially usable. The v0.2.0 line relicenses to Apache 2.0, swaps the BundleFusion pose backend for a GPU Theseus optimizer, adds an ONNX FoundationStereo depth model, USD/USDZ export beside textured OBJ, optional cross-view color fusion for uneven lighting, and ARM64 / Jetson Orin support.
The headline is not a feature, it's the license. A production-grade reconstruction pipeline from NVIDIA that you can actually ship in a commercial product, with USD export so it drops straight into an Omniverse or simulation stack. Sub-millimeter on small objects from calibrated stereo, roughly half an hour on a single workstation GPU. The catch worth respecting is the input: this wants calibrated synchronized stereo, not a phone video, so it is a rig tool, not a casual-capture tool.
Physical AI agent skills: agents that automate the physical-AI workflow, including turning fleet-captured video into editable 3D scenes for simulation and synthetic data, built on Cosmos 3 with Isaac Sim and Omniverse
The reconstruction step is being absorbed into an agent-driven pipeline. Agents handle scene preparation, data capture, simulation control and validation, and one skill turns real fleet video into editable 3D scenes you can re-light, re-arrange and re-simulate. For 3D people the message is direct: the manual middle of the pipeline is automating, and the value moves to the judgment around it.
I keep coming back to the Jacquard loom. Once punched cards drove the threads, the weaver stopped pulling them and started designing the pattern instead. This is that, for our pipeline. The agent does the scene prep and the capture; the reconstruction becomes a step it runs, not a craft you hand-tune. The honest read: this is NVIDIA's stack talking to NVIDIA's stack, Cosmos plus Omniverse plus Isaac, so the lock-in is real. But the direction is not NVIDIA-specific, and the people who stay valuable are the ones who own the judgment the agent cannot: what to capture, what good looks like, where the reconstruction lies.
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.
