sceneforge results

2 new experiment reports · 3 earlier captures · published September 10, 2026

Latest real2sim experiments

Independent assets · UR5 CB3
Independent assets · UR5 CB3
Three scans · 28 table + 28 pedestal + 51 final-scene views
Latest CB3 video · independent assets · comparison images
47-view laboratory · verified bottle pick
47-view laboratory · verified bottle pick
47 D455 RGB-D views · reconstructed laboratory · UR5e + Robotiq
Pick + wrist videos · held-out comparisons · verification

Selected videos, renders and verification results from the two latest experiments. Each report uses its own capture and evaluation setup; scores are not directly comparable across experiments.

How this works

90 seconds, end to end: 58 photographs, camera poses, metric depth, segmentation, then two branches — the background reconstructed as 3D Gaussians with the objects carved out, and the objects rebuilt as meshes — recombined in MuJoCo. Every 3D frame is real pipeline output; every number on screen is measured.

Captures

47-frame depth
47 framesdepthaccum 16
sceneforge pipeline · VGGT-1B (feed-forward)
session_20260902_184520
58-frame ir
58 framesir
sceneforge pipeline x2 · splat trainer ablations x6 · asset extraction candidates x4 · end-to-end demonstrations x9 · VGGT-1B (feed-forward)
session_20260902_171546
8-frame
sceneforge pipeline x3 · VGGT-1B (feed-forward) · baselines
session_20260901_180746

Each capture page compares every method run on that same footage. Numbers are only comparable within a capture.

Interactive

Splat viewer — 34 trained background(s), free-flown in the browser and compared two at a time under one camera. Renders client-side, so it needs no GPU and nothing running on the cluster. Videos below are fixed paths; this is not.