3D Intelligence Report – June 12, 2026
Five verified finds. Today's thread is efficiency and infrastructure: the paper shrinks world models to laptop scale, the tool prunes splats without rendering, NVIDIA pays top dollar to maintain a free library, and Leica rebuilt its scanner line around real-time pipelines.
Every link below was fetched and verified on June 12, 2026, the day this report went out.
Light-WAM: Efficient World Action Models with State-Fusion Action Decoding
Ziang Li, Dongzhou Cheng, Yibin Wang, Shiyue Wang, Xiaoyang Xu, Lingxuan Weng, Juan Wang, Jiaqi Wang
0.44B params, 72 ms, 4.1 GiB
World action models without the data-center bill: state-fusion action decoding reads adapted states from multiple backbone layers and predicts action chunks in one pass, no heavy generative action module. 0.44B trainable parameters, 72 ms inference, 4.1 GiB peak GPU memory, with strong LIBERO results and multi-task RoboTwin 2.0 performance. Full codebase released (src, configs, experiments, eval logs).
Every world-model paper this month chased capability; this one chases the electricity bill. One forward pass instead of a generative action module, and the whole thing trains and runs in 4.1 GiB. That number matters more than the benchmark: it's the difference between a lab capability and something a student or a small robotics shop can actually fine-tune. Efficiency is how this field stops being a spectator sport.
Senior System Software Engineer – Neural Graphics Performance
$152K-$287.5K base (L3-L4) + equity
The posting says it in NVIDIA's own words: 'Love GSplat? So do we.' The job is pushing gsplat, the open-source splatting library half this community trains with, plus the Omniverse NuRec SDK, to full hardware speed with CUDA and Slang. A trillion-dollar company paying up to $287K base to accelerate a free library means that library has become infrastructure. A merged performance PR in gsplat might be the strongest application you can send.
Four funding doors, four regions
next deadline in 6 days
-
Asia deadline 2026-06-18 · 6 days leftSIGGRAPH Asia 2026 (Kuala Lumpur): XR, Technical Communications, Real-Time Live
venue: Dec 1-4, 2026; XR due Jun 18, Tech Communications Jul 27, Real-Time Live Aug 7
The XR track closes in six days, and it's the most forgiving door into a SIGGRAPH venue: demos, not full papers. Technical Communications (4 pages, July 27) is the right size for that solid-but-small result sitting in your drawer. Kuala Lumpur in December is a cheap fly-in compared to the US venues.
-
Europe deadline 2026-06-30 · 18 days leftCopernicus Masters 2026 (ESA / AZO Earth observation competition)
prize pool above EUR 1.5M across challenges, incl. a small satellite mission
The one EO competition where a working prototype beats a polished proposal. If you've built anything that turns Sentinel data into decisions (3D change detection, urban twins, vegetation structure), eighteen days is enough to package what already runs. Note: the dedicated competition site served an expired certificate today; the EU portal link is the safe door.
-
Americas deadline 2026-07-27 · 45 days leftNSF SBIR/STTR Phase I (NSF 26-510, + 26-511 instrumentation pilot)
up to $305K Phase I (to $2M cumulative), zero equity; Project Pitch required first
The program is back after its pause, with $250M behind it and a new instrumentation lane that fits lidar and 3D capture builders unusually well. The trap is sequencing: you need an invited Project Pitch before the full proposal, so the real deadline is around June 20, not July 27. Zero equity taken is still the best deal in American deep-tech funding.
-
France deadline 2026-09-07 · 87 days leftGENCI / Jean Zay: national HPC-AI hours, Regular Access session
national GPU allocations (1 normalized h = 0.25 H100 h), free, 1-year validity; smaller Dynamic Access requests rolling
Funding nobody thinks of as funding: compute. A Jean Zay allocation pays the part of a 3D deep learning project that actually hurts, and the Dynamic Access lane (up to 50k normalized GPU hours) answers in weeks, year-round. If you're France-affiliated and still renting cloud GPUs for research, you're leaving public money on the table.
REFINE initial release
initial release
A plug-and-play, rendering-free pruning script for any pre-trained 3DGS model: an analytically approximated Hessian field scores every Gaussian's importance without a single forward render, cutting pruning-related compute by ~3,000x versus rendering-based methods at competitive quality. One Python file, deps: torch, numpy, plyfile.
The whole tool is one script you point at a trained .ply, and it needs no renderer to decide which Gaussians earn their memory. With splats becoming an OS primitive and an interchange standard, pruning is no longer an academic nicety, it's what makes your scenes shippable. Run it on a scene you know; watching which 40 percent of your Gaussians were dead weight is an education in itself.
Leica RTC series launch: RTC300/500/700 terrestrial laser scanners with real-time field-to-office streaming
Up to 2,000,000 points per second, ranges to 270 m, and one platform replacing two long-running families (RTC360/LT and the ScanStation P-series). The headline shift is workflow: scans stream to the office in real time instead of riding home on a USB stick. Capture hardware is converging on the same lesson software learned: the value is in the pipeline, not the device.
The spec everyone will quote is 2 million points per second; the spec that changes daily work is the streaming. When the office sees your scan while you're still on site, registration problems get caught before you pack the tripod, and the second site visit (the most expensive line item in scanning) starts disappearing. Hardware consolidation also tells you where margins are going: one platform, many price points, software on top.
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.
