Coffee workspace — recorded gravity simulation

Each clip plays five seconds of simulation at 1x speed and 30 fps. RGB and clay
show the same measured rigid-body trajectory, with a fixed camera for each
method. The scenes use the page's edited reference image with a plain white mug
and the corresponding refreshed mug visuals. Saved scene layouts are preserved.

SceneRig: the teaser's GPT-6 Astra reconstruction, with its authored PhysX
collision proxies and physical properties. Physics runs at 60 Hz; recorded poses
are rendered in Blender from the existing teaser camera. The RGB/clay material
change does not alter geometry or motion.

VIGA: the page's Claude Opus 5 object-grouped reconstruction, with one dynamic
body per object group. The desk, wall and whiteboard are static. Gravity is aligned
to the desk; static surfaces use their original triangle geometry, avoiding an
enclosing convex hull across the desk's open spaces. Dynamic parts use the
existing grouped harness's convex hulls and 1 m/s depenetration cap. Physics runs
at 60 Hz. This recording adapter does not alter the paper's reported scores.

SimFoundry: the saved scene in native OmniGibson / Isaac Sim, with its authored
collisions and physical properties, at 120 Hz. Objects are woken from the saved
sleep state without impulses or pose changes. The existing white support visualization,
lighting and recovered camera are retained; the robot is hidden. Clay is rendered
by replaying the RGB recording's saved poses without stepping physics again.

SceneRig and SimFoundry retain their original collision proxies after the mug
visual refresh. Methods have independently reconstructed geometry and cameras;
their views are not identical camera calibrations. A stable scene may appear
still while the playback clock advances. Loops reset to the initial state.

These recordings illustrate this scene; they are not new aggregate benchmark
measurements. No motion is synthesized or interpolated between endpoints.