Bedroom find: the model identifies floor clutter beside the bed.
Explore
Recorded runs and selected 3D worlds
Watch a model inspect a synthetic room, then open a selected world yourself. Saved replays show where it turned, what it flagged, and what it missed. They illustrate specific runs, not performance across the full photo benchmark.
01 · Recorded runs
Watch recorded runs
These are recordings, not a live model service. Open a trace to see the model's actions beside the intended hazard.
Bathroom miss: the model looks toward a loose bath mat but flags support fixtures instead.
Staircase loop: the model repeatedly flags an off-target spot and misses the seeded stair hazard.
02 · Selected worlds · 11
Walk a selected world
Open a browser-based 3D room to inspect a hazard from another angle. These are selected reconstructions based on synthetic scenes, not identical copies of every benchmark image. The photo leaderboard is scored separately.
03 · Walk mode
Step-by-step traces
Where a saved model trace exists, open it for that world to compare the model's actions with the seeded target. Traces cover 4 of the 11 selected worlds.
Want to inspect the underlying evidence? Start with the results or browse the code and data.