3D Intelligence Report – September 5, 2026
Theme of the day: five separately-owned pieces of the same stack, all moving to fit together. A paper stops treating physics, geometry and appearance as three separate models and reads them off one camera. A drone maker hires someone whose whole job is turning raw flight logs into a dataset the rest of the team can trust. An open-source splat renderer drops its NVIDIA-only requirement and runs on AMD instead. A survey-instrument company buys the sensor that complements the one it already sells. And France's one call that funds almost anything opens its window, generic edges and all.
Every link below was fetched and verified on September 5, 2026, the day this report went out.
Puffin-World: Scaling a Unified Multimodal Model with Native 3D World States
Kang Liao, Yihang Luo, Xiao-Ming Wu, Linyi Jin, Size Wu, Chunyu Lin, Yao Zhao, Fei Wang, Wei Li, Chen Change Loy
12 of 12 best median-error results and 33 of 36 best AUC metrics across camera-understanding and generation benchmarks; 0.26 degree roll error on LaMAR, 17.22 PSNR on RealEstate10K 3D view synthesis
Puffin-World is the second release in the Puffin line, and it scales the same idea up: read a camera the way a physicist, a surveyor and a painter would read it at once. It names three native world states inside one model. Physics is gravity-aware camera understanding, so the model tracks how a shot behaves under real motion instead of treating every frame as weightless. Geometry is dense 3D structure pulled straight out of the same forward pass, no separate reconstruction stage. Appearance is what stays visually coherent as the camera moves. All three train together on Puffin-16M: 15 million vision-language-camera triplets, a million camera trajectories, 44.5 million camera-labeled images pooled from 28 public datasets, weighted toward the pitches, yaws, rolls and compound moves that break most camera-pose models.
What gets me about this one is which problem they decided was actually the same problem. Most pipelines I've built keep camera pose, 3D structure and rendered appearance as three separate stages talking to each other through files on disk, and every handoff is a place error creeps in. Puffin-World trains all three off the same backbone instead, and the physics piece is the one I did not expect: gravity-aware trajectory understanding, so the model has an opinion about which way is down before it has an opinion about anything else. That is basically what your inner ear does before your eyes catch up. The Puffin-16M weighting toward hard compound camera motion is the unglamorous part that will actually matter if you try to use this on real handheld or drone footage instead of a lab dataset. Code, weights and the dataset are all up, which is rare enough at three days old that I checked twice.
Senior Autonomy Engineer, Data Curation
$170,000-$240,000 (disclosed)
This title would have gotten shrugged off two years ago as a data-ops role, and I think that undersells it badly. Every 3D pipeline I have run lives or dies on what goes in before the model ever sees it, and curating flight logs from a drone that is actually flying in the real world, wind, glare, birds, other drones, is a much harder curation problem than curating a static photo set. Skydio splitting this out as its own senior, well-paid role tells you they have learned the same lesson the hard way: a model is only as honest as the mess it was trained on. If you have ever spent a weekend throwing out half your scan captures because the drift made them useless, you already know what this job actually does all day.
ANR AAPG 2027 (Appel a Projets Generique), step 1 pre-proposal
not published on the call page itself; this is ANR's annual generic call and covers its standard project grant lines (PRC, PRCE, JCJC among them), the amount is set per project during instruction, not published as a headline figure
I will say the honest thing about this one: it is not a 3D-specific call, it is THE call, France's one annual generic line that funds nearly everything academic research does here, spatial AI included. That is also exactly why it belongs in this pack. If you are a France-based researcher or a small lab and you have been waiting for a dedicated point-cloud or spatial-AI funding line to show up, stop waiting. Your project almost certainly fits under one of the 57 themes, and the JCJC line specifically exists to fund early-career researchers running their first independent project. Thirty-eight days is enough runway to get a pre-proposal in if you already have the idea written down somewhere. The catch: no domain-specific evaluators means you have to make the spatial-AI case yourself, in plain language, to a generalist panel.
vkSplatting (vk_gaussian_splatting) 2026.2.9
the same Vulkan ray-traced splat renderer now runs on AMD hardware, no NVIDIA-only SDK required
vkSplatting is NVIDIA's open testbed for real-time 3D Gaussian Splatting rendering over Vulkan: rasterization, ray tracing and hybrid paths in one repo, 494 stars, 56 commits, real source under src/shaders/cmake, not just prebuilt binaries. 2026.2.9 trims the per-ray payload in the stochastic ray-traced path by around 8 data words per ray, fixes a depth-arbitration bug that made multi-material splats disappear in raster and hybrid modes, and, this is the headline, adds AMD RADV driver support. The ray-traced renderer no longer needs NVIDIA's proprietary NGX SDK to build and run: it now builds and runs on Radeon GPUs over the open RADV driver.
I have watched splat rendering treat the GPU vendor as part of the spec for two years now, half the fast viewers assume you're on an NVIDIA card and the other half don't ray trace at all. NVIDIA open-sourcing the fix to its own lock-in is the part worth sitting with. RADV is the open-source AMD driver, not a vendor SDK, so the dependency they removed was one they controlled. That is a company betting the API (Vulkan) matters more than the hardware it ships. Practically: if you've been holding off on ray-traced splat viewing because you didn't want to be tied to one card, this is a testbed, not a shipping product, but it is the first time I've seen the actual renderer code demonstrate it is possible.
Agreement to acquire Guidance Marine, a UK maritime relative-positioning sensor maker, to pair with Hexagon's existing absolute GNSS positioning line
Guidance Marine, founded 1991 and based in Leicester, builds radar, laser, vision and microwave sensors that measure a vessel's position relative to nearby structures, offshore platforms, wind turbines, quaysides, other vessels, and feeds that into Dynamic Positioning systems. Hexagon sells absolute positioning through GNSS. Put them together and a vessel gets both where it is on the globe and how close it is to the thing it must not hit, in the same stack. Completion is expected by year-end 2026. Named target markets: offshore wind, marine construction, port automation, vessel autonomy.
Absolute position and relative position are answers to two different questions, where am I on Earth versus what am I about to hit, and most navigation stacks I've seen bolt them together from two vendors and hope the coordinate frames agree. Hexagon just bought the second half of that answer outright instead of integrating around it. The revenue number is the tell: above-average margin on a EUR 22 million business says Guidance Marine wasn't a rescue, it was an acquisition made from a position of strength on both sides. Watch where the next fix-and-relative pairing gets bought next: aerial and ground robotics have the exact same two-question problem, and right now most of them are still solving it with two separate vendors too.
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.
