Dvidia/piper-rgbd-tracker-v01

DVIDIA · Known-block RGB-D tracker v0.1

A small CPU estimator for one known upright orange 40 mm cube in native rendered RGB-D. This is an evidence archive, not a pretrained robot policy or physical qualification.

Resolution	Clear estimates	Position error p95	Yaw error p95	Estimator wall p95	Registered gate
160×120	12/16	2.34 mm	11.31°	7.73 ms	Failed coverage and yaw
320×240	16/16	1.98 mm	6.36°	7.58 ms	Passed clear-scene gate

All 256 negative fixtures withheld estimates. Partial occlusion returned 0/32 estimates and remains unsupported. Ten conditions and two resolutions share 16 pose groups, not 320 independent samples. Error percentiles are conditional on estimates being available. Uncertainty values are uncalibrated geometric scales. Timing is offline shared-host compute, not sensor latency or real-time qualification.

Read or run

Interactive images and results · Research · Source and commands · GitHub preview

Download piper-rgbd-tracker-v01.zip for all actual RGB/depth arrays, immutable fixture inputs, source freeze, prediction receipts, truth-separated scoring, report, independent audit and development history. inventory.json binds every archived evidence file. The nested study JSON is not a tabular dataset-viewer configuration.

The tracker needs Python 3.12 and NumPy; generating fixtures also needs MuJoCo 3.15 and a working OpenGL context. macOS arm64/CGL is the evaluated renderer. No hosted model/API key or dedicated GPU is needed for estimation; minimum hardware and other rendering backends remain unqualified. Install dependencies/assets once, then use locally. No motor-learning export was enabled.

Evidence and limits

Runtime source freeze 846a1d9b04745a516ff060b4efa931e5e9898fbf. One development-only brightness change 65/20→25/8 preserved geometry/occlusion checks; confirmation ran once with no outcome-based tuning. Independent NumPy audit recomputed 320 rows, 20 summaries and 2 gates with zero mismatches, checked 2,578 evidence files and 32 occluder-depth samples. Audit SHA-256: a9168125d0bc233385ad07e22414ccc447557a0becc044ab4c2597dcb93fe840.

Static authored rendering poses include elevated unsupported objects; no manipulation physics rollout or motion-learning claim. Occlusion cards/duplicate cubes are explicit render-only challenge geometry. CGL depth accuracy remains unverified. Cube yaw is identifiable modulo 90°. Tilted objects, partial occlusion, arbitrary assets, human-video action recovery, hardware and arbitrary-arm transfer remain unsupported.

Original DVIDIA source/generated records MIT. Upstream PiPER/MuJoCo notices are included; no third-party meshes, weights, private recordings or credentials are bundled.