3D Intelligence Report – October 22, 2026
Theme of the day: points that persist through time. A world model forecasts future video and the 3D path of each scene point. World Labs pays up to $325K base to render synthetic worlds at fleet scale. Rerun 0.40 lets an agent read a recording's transform graph. ICC 2027 opens its Warsaw call around big data cartography.
Every link below was fetched and verified on October 22, 2026, the day this report went out.
Learning the 4D World through Point Trajectory Forecasting
Byungwoo Jeon, Sunghwan Hong, …, Hanbyul Joo, Marc Pollefeys, Saining Xie, Jinwoo Shin (KAIST, ETH Zurich, SNU, NYU, RLWRLD, Physics Sim Lab)
1,200 multi-view clips with 3D point tracks
Video world models predict pixels, and pixels forget which point is which. The Video-Point Model (VPM) takes multi-view video history and a language instruction and forecasts two things at once: future videos from arbitrary viewpoints, and the 3D trajectories of persistent scene points, where each trajectory follows one point across all views. It extends a pretrained world model (NVIDIA Cosmos3-Nano) with camera-conditioned multi-view processing and a trajectory head. On the new VP-WorldBench (1,200 synchronized multi-view clips with 3D point tracks) and on existing 4D generation benchmarks, the authors report that point supervision improves both video fidelity and 3D motion accuracy over a video-only counterpart. Adding robot actions as a token modality raises manipulation success in simulated and real-robot experiments. Authors include Marc Pollefeys, Saining Xie and Hanbyul Joo. 42 upvotes on Hugging Face papers.
What I like is the choice of target. A video model can paint a convincing next frame and still lose track of which point is which, and anyone who has matched point clouds across epochs knows that identity is the whole problem. VPM asks the model to keep each point's 3D path, and the authors report the video gets better for it, not worse. That is a strong hint that geometry is a training signal, not just an output. My caution: it is a non-commercial licence, and the gains are reported against their own video-only baseline, so I would wait for outside numbers before building on it.
Rendering Systems Engineer, Simulation and Synthetic Data
$250K to $325K base, plus equity
This posting tells you where world-model training data comes from: a game engine run thousands of times in parallel. The interesting part is that the hard problem is plumbing, not art. Many worlds, many cameras, GPUs kept busy, I/O kept out of the way. If you came to 3D through graphics and performance work rather than research, this is the kind of role that values you. One honest note: the bar is Unreal source-level renderer work, which few people have done, so read the qualifications before you apply. World Labs has appeared several times in this report as a company, and this is the first time as an employer.
ICC 2027 Call for Papers (33rd International Cartographic Conference, Warsaw)
Not a grant. A call for papers
Big data cartography is a theme a point cloud person can walk into without changing a word of their work. Billions of points are spatial big data, and the question cartographers ask, how do you show this to a human so they understand it, is exactly where most 3D projects stall. I would bring the visualization side: level of detail, classification you can read, a map product from a scan. Full papers close November 20, abstracts December 1. If your experiment is done, the paper is the better choice. Check the topics page before you write.
Rerun 0.40.0
Agents read the transform graph over MCP
Published 2026-10-21 at 10:25 UTC, not a pre-release, 19 assets attached. From the release notes: a viewer design refresh with a new home screen, a command palette in the top bar and consistent typography and colours; agents can inspect transform graphs through MCP with rerun_get_transform_graph; catalog API improvements with Segment and Layer handles, DatasetEntry.schema() and DatasetEntry.info(); the selection panel shows which layers a component comes from; asset coverage editable in the viewer; Asset3D meshes without normals are now shaded with computed flat normals. Experimental: table filters and sorting in the table blueprint, LeRobot dataset metadata via LeRobotReader.metadata(). Breaking: Python 3.11 is the minimum supported version. Install: pip install -U rerun-sdk.
The line I would test first is the MCP tool. A wrong transform is the most common silent bug in any capture rig, a lidar mounted ten centimetres off or a camera frame flipped, and now an agent can read the whole transform graph and tell you where things sit relative to each other. That is a debugging assistant with a real view of your scene, not a guess from your code. The shading fix matters too, a mesh without normals no longer looks like a flat silhouette. Check the Python 3.11 minimum before you upgrade an older environment. I covered 0.36 in August, and the direction since then is clearly toward agents and datasets.
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.
