Skip to content

Latest commit

 

History

History
40 lines (30 loc) · 1.67 KB

File metadata and controls

40 lines (30 loc) · 1.67 KB

Metric3D clearance field A0.1 normal-guided successor

Date: 2026-08-03

Status: FROZEN_BEFORE_NORMAL_GUIDED_WALKING_RPY_OUTCOME

Single changed variable

A0 failed under fresh roll/pitch/yaw motion with a depth-only ground plane fit. A0.1 keeps the depth model, point cloud, bands, obstacle-height interval, clearance percentile, collision horizons, unknown semantics, and all seven A0 fresh gates unchanged. It changes only ground-frame estimation.

The official Metric3D output prediction_normal contains a camera-frame unit surface normal and a positive kappa concentration per pixel. A0.1:

  1. uses pixels in the lower 45% of the image;
  2. keeps normals with absolute camera-y component at least 0.55;
  3. keeps the upper half by per-frame kappa median;
  4. sign-aligns candidate normals toward camera-up and takes a capped-kappa weighted consensus;
  5. forms 4 cm bins of point-to-plane offset in the accepted 0.45--2.20 m camera-height interval;
  6. refits the densest support with SVD using an 8 cm offset neighbourhood and a 15 degree normal-consistency limit.

No IMU, registered depth, camera pose, person box, class, or future frame is an input to the candidate.

Data roles and decision

  • Development mechanism screen: consumed walking_rpy, exact frozen 180 RGB frames from A0.
  • If and only if all original seven A0 gates pass, freeze and evaluate the official walking_halfsphere sequence as new fresh evidence.
  • Otherwise stop this normal-guided successor. Do not tune normal thresholds, kappa selection, offset bins, or refit limits on walking_rpy.

This successor does not reopen the A0 fresh terminal or authorize A1, Android, reminders, or safety claims.