3D Intelligence Report – September 21, 2026
Theme of the day: reality capture going production-grade, from three directions at once. A wearable scanner claims native 3 mm point clouds while you walk, the detail tier that used to mean a tripod. NVIDIA gives its splat reconstruction its own open repo and exports to the OpenUSD splat standard, which is what a pipeline component looks like. Zoox pays up to 244,000 dollars for someone to render sensor data with Gaussian splats and prove it matches the real car, which puts splats inside a safety process. And a Copernicus and Galileo hackathon opens registration on September 28.
Every link below was fetched and verified on September 21, 2026, the day this report went out.
Machine Learning Engineer, 3D Sensor Simulation
USD 173,000 to 244,000 base, disclosed on the posting
This role puts Gaussian splatting inside a safety process, and the realism metrics are the line I'd underline. Rendering a pretty splat is the easy half; the job is to say, with a number, how far the synthetic LiDAR return sits from the real one, and to defend that number in front of a safety team. That moves Gaussian splatting out of content creation and into verification, where a wrong frame is a test that lied to you. It's the third AV company in this pack's history hiring for neural rendering, after Torc in June and Applied Intuition in July, and the job descriptions keep drifting from 'make it look real' toward 'prove it's real enough'. If you've been building splats for visuals, the skill worth adding is measurement: camera models, sensor noise, and a metric you'd stake a release on.
12th CASSINI Hackathon, Space for Peace and Resilience (registration opens Sept 28)
Finals: EUR 5,000 / 3,000 / 1,000, plus 6 months of mentoring for the top 3. Local sites add their own sponsor prizes.
Hackathon prize money is never the point, and EUR 5,000 split across a team won't change anyone's year. The six months of mentoring and a demo in front of EUSPA might. What I like about this one is the data: Copernicus imagery, Galileo positioning and EGNOS corrections are free, and plenty of people in 3D have never touched them because nobody forced a weekend on it. Resilience is a topic where a point cloud, a change map and a precise position actually do something, flood extents, damaged structures, access routes. Registration only opens on September 28, so today is for finding two teammates, not for clicking.
fVDB Reality Capture v0.6.0
Splats out as OpenUSD, Apache-2.0
The Gaussian splatting APIs move out of fvdb-core into this new repo, so the low-level CUDA kernels and the high-level reconstruction API now live apart. v0.6.0 adds multi-GPU 3DGUT support and fixes a multi-GPU backward-rasterisation race, releases GARfVDB, adds dense depth supervision (DepthMapAttribute), scale regularisation and Gaussian isocontour tile intersection, updates USD export to the OpenUSD ParticleField3DGaussianSplat standard, fixes NuRec USDZ export at SH degree 1 and 2, adds a checkpoint API with recovery, and supports PyTorch 2.13+. No benchmark numbers were published with the release. To try it: clone the repo and pip install -e . on a CUDA machine with PyTorch 2.13+, with fvdb-core as the dependency (the README has the steps).
The feature list is long, but the change I care about is structural: kernels in one repo, reconstruction in another. That's the moment a research codebase admits it has users who want to call it without reading it. The OpenUSD export matters for the same reason. A splat that lands as a standard USD particle field can go into a scene pipeline that never heard of fVDB, and checkpoint recovery means a long multi-GPU run survives a crash. None of that makes a prettier splat. All of it makes a splat you can schedule on Monday and still trust on Friday. I'd test the depth supervision first, since that's where captures with thin or shiny surfaces usually fall apart.
The VLX 4, a wearable scanner pitched as the first to reach native 3 mm point cloud resolution while you walk
Walking capture is moving into the detail tier that used to require tripod terrestrial scanners: LOD350+, Scan-to-BIM and MEP-grade elements such as thin pipes, bolts, brackets and sharp edges. The VLX 4 carries a three-LiDAR array including Hesai's new MT LiDAR, which NavVis describes as the first LiDAR purpose-built for reality capture and co-developed with them. Launched around INTERGEO 2026. NavVis has not disclosed price, availability date, or speed and range figures.
The number is 3 mm, but the claim that matters is 'while walking'. For years the deal was simple: walk for coverage, set up a tripod for detail, and stitch the two. If a wearable really holds 3 mm on a bolt or a thin pipe, that trade shrinks, and the tripod becomes the tool you bring for control points and edge cases. I'd hold off celebrating until someone publishes a side-by-side against a terrestrial scan on the same MEP room, because 'native resolution' and 'accuracy' aren't the same word. The other signal is the sensor: a LiDAR maker co-designing a unit for reality capture tells you this market is now big enough to get its own hardware.
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.
