Skip to content

Add a local RGB-D block tracker with held-out native-camera evidence - #31

Draft
Tarzelf wants to merge 2 commits into
feat/piper-incremental-hold-v03from
feat/piper-rgbd-tracker-v01
Draft

Tarzelf wants to merge 2 commits into
feat/piper-incremental-hold-v03from
feat/piper-rgbd-tracker-v01

Conversation

@Tarzelf

@Tarzelf Tarzelf commented Nov 1, 2026

Copy link
Copy Markdown
Contributor

Add a causal CPU RGB-D baseline that estimates one known upright 40 mm orange cube from image/depth/calibration, with explicit abstention and four provenance-bound pose/uncertainty signals. The existing action recorder and learning export remain unchanged.

A frozen, once-only confirmation uses 16 held-out pose groups × 10 conditions × 2 resolutions. At 320×240, clear coverage is 16/16, position-error p95 1.98 mm, yaw-error p95 6.36° and estimator wall p95 7.58 ms. At 160×120, coverage and yaw gates fail. All 256 negative fixtures abstain; partial occlusion returns no estimates in 32 cases and remains unsupported. These are correlated static simulation fixtures, not a learned skill or hardware qualification.

Validation: 34 tracker/runner tests, 18 independent-audit/packaging tests, 37 capture/action-contract regressions, 6 publication tests; independent saved-evidence audit of 320 rows/20 summaries/2 gates with zero mismatches and 32 occluder-depth checks. Real browser verifies result selection, missing-object abstention, desktop and 444 CSS-pixel layout. Source was frozen before confirmation; original development and graphics failure records are retained.

This draft targets the preceding research preview branch. It does not promote stable main, send physical commands, or enable motor-learning exports.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant