From pixels to position.
A small local estimator looks for one known block in color and depth. This experiment measures when it can estimate a pose—and when it should say it cannot.
Static rendered fixtures. One orange, upright, 40 mm cube. No new learned manipulation, task success, or hardware qualification. “Available” means the estimator returned a pose; it does not mean that pose is correct.
All confirmation conditions
160 px: gate failed; 320 px: clear-scene gate passed. Partial occlusion is reported separately and remains unsupported.
Each row contains 16 held-out pose groups. Variants and resolutions from one group are correlated; they are not extra independent examples. Error percentiles use available estimates only, so read them alongside the coverage.
What this measures
RGB and depth come from the same paused native renderer. The estimator receives pixels, optical-Z depth, declared camera calibration and timing metadata. Native object pose enters only the offline evaluation after predictions are saved. Occlusion cards and duplicate objects are explicit render-only challenge geometry; these are perception fixtures, not physical manipulation trials.
Cube yaw is scored modulo 90°. Tilted objects are outside the estimator’s upright profile. Uncertainty values are geometric heuristics, not calibrated confidence bounds. Synthetic stale/future timestamps test the data contract; measured compute time does not establish sensor delivery latency or a real-time robot deadline. The renderer’s depth-accuracy limitation remains explicit.