MOTION_CONDITIONED_OCCUPANCY_A0.md

Motion-conditioned occupancy A0

Date: 2026-08-03

Status: FROZEN_BEFORE_WINDOW_LOSO_OUTCOME

Question

Can HFTF-style causal camera-motion evidence make a low-capacity collision occupancy probability transfer across whole windows, where depth-only and confidence-stratified post-processing failed?

Fixed data and isolation

• Inputs: fixed-world-floor UniDepth reports for the four consumed walking_static/xyz windows and six consumed walking_rpy windows.
• Evaluation: leave one complete 3-second window out; all band×horizon rows from that window are held out together.
• Labels: registered-depth fixed-world-floor occupancy, read only by the fit and evaluator for the appropriate folds.
• No fresh source is opened for this Development screen.

Fixed features

Per band×horizon row:

• predicted clearance minus horizon;
• predicted clearance and horizon;
• UniDepth clearance-support log1p confidence;
• ground-plane median residual;
• log1p obstacle point count;
• left and centre indicators (right is reference);
• causal RAFT-small previous-to-current global median/P90 flow magnitude;
• normalized affine translation x/y, absolute rotation and log scale;
• affine inlier fraction;
• residual-flow median/P90 magnitude;
• first-frame motion-missing indicator.

RAFT-small uses the already frozen official checkpoint SHA-256 01064c6dba73b0fc9fc8edf772248560a00a3acfd62ac6677e9eeebad9680e27 at 224x128. The classifier is standardized unweighted Logistic Regression, fitted deterministically with L2 coefficient 0.01; no polynomial features, class weights, threshold search, or hyperparameter selection are allowed.

Continuation gates

The same six probability gates remain:

1. Brier reduction >=15% versus deterministic clearance;
2. log-loss reduction >=20%;
3. ECE <=0.10;
4. false-clear among P(occupied)<=0.05 outputs <=5%;
5. such high-confidence-clear coverage >=10%;
6. occupied recall at P(occupied)>=0.50 >=85%.

All must pass in pooled window-LOSO predictions before a final model may be frozen for walking_halfsphere.