DVIDIA known-block RGB-D tracker v0.1 · research preview

A local known-block RGB-D estimator and reproducible static native-camera benchmark.

• 320×240: 16/16 held-out clear estimates; position-error p95 1.98 mm, yaw-error p95 6.36°, estimator wall p95 7.58 ms.
• 160×120 fails the frozen coverage/yaw gate. All 256 negative fixtures abstain; partial occlusion remains unsupported (0/32 estimates).
• Independent audit: zero mismatches, 320 cases, 2,578 evidence files, 32 occluder-depth checks.

Sixteen pose groups generate correlated variants; this is not 320 independent trials. Static known-object perception only. No learned policy, physical qualification or action-export integration. GPU-free estimation; rendering separately requires a working OpenGL context. Depth accuracy and uncertainty scales remain uncalibrated.

Research and reader · Hugging Face evidence · Local reproduction

Evidence archive: 49,451,665 bytes; SHA-256 9676e71b3a0074bd4fe1b19a8e77a08d7366fa71feb3ba041aa633af6eb60f33. Original DVIDIA source/generated records MIT; upstream notices retained. No private footage, credentials, meshes or model weights bundled. Research preview; not a replacement for the stable application.