Buckets:

AsianPlayer
/

TriDrive

Files

xet

AsianPlayer/TriDrive / benchmark /tables /paper_tables_iclr /tab_future_state.tex

AsianPlayer

28 days ago

download

raw

6.65 kB

% tab_future_state.tex --- GENERATED by paper/tables/gen_tables.py --style iclr on 2026-09-18; do not edit by hand.
% Every number is copied verbatim from these sources (path relative to the repository root : section / line):
%   - results/tables/future_state_eval.md : Verdict counts (table line 156)
%   - results/tables/future_state_eval.md : Occupancy BRIER, engaged head_turned 3s rows (table line 18)
%   - results/tables/future_state_eval.md : Occupancy AUROC, engaged head_turned 3s rows (table line 87)
%   - results/tables/future_state_eval.md : Recovery time / first-leave MAE (table line 209)
% Full caption of the IEEE draft (provenance, verdict legend, line pointers; kept as a comment):
%   | F2$'$ future-state head: a detached horizon-conditioned head ($\approx$0.034\,M params) on the FROZEN joint model's driver state $z_U$ + label-free
%   | dwell features, trained on the train anchors of both regimes, selected on validation occupancy BCE; test split, seeds 0 / 1 / 2, device-cluster
%   | bootstrap ($n_{\mathrm{boot}}{=}500$). States hands\_on / head\_turned (v2 causal kin targets) and gaze\_off (own landmark chain, validation pending).
%   | Baselines: persistence $\pi_k = z_0$; dwell table = train empirical $P(s_{t+k} \mid z_0, \text{time-in-state bin})$; FINAL closed form = forward
%   | marginal of the frozen model's own hazards (hands / head only). (A) Verdict counts over the cells regime $\times$ state $\times$ horizon (1 / 2 / 3 /
%   | 5\,s); (B) engaged head\_turned at 3\,s; (C) MAE in s of $E[\min(T, 5.2\,\mathrm{s})]$ for the first-leave time (all anchors) and the remaining dwell
%   | in the off-task state, paired $\Delta$ per seed and CIs in the source. $\Delta$ = head $-$ baseline; Brier / MAE lower better, AUROC higher. Verdicts
%   | follow \texttt{kill\_checks.beats} ($\geq 2/3$ seeds CI excl.\ 0, same sign, $|$mean $\Delta| >$ seed sd): better / worse = significantly better /
%   | worse; n.s.\ (INCONCLUSIVE) and not worse (NOT\_WORSE) = no significant difference. Nothing here changes the headline model. Sources:
%   | future\_state\_eval.md verdict counts (line 156), Brier (line 18), AUROC (line 87), recovery time (line 209).
\begin{cwtable*}{tab:future_state}{Detached future-state head on the frozen \cwname{} driver state (BATON test, 3 seeds): occupancy of hands-on / head-turned / gaze-off at 1--5\,s against persistence, a label-free dwell table and the model's own closed form. (A) Verdict counts over cells; (B) engaged head-turned at 3\,s; (C) recovery-time MAE in s. Better = significantly lower Brier or MAE, higher AUROC.\label{tab:tab_future_state}}{l l r r r r r}
\multicolumn{7}{@{}p{0.98\textwidth}@{}}{\textbf{(A) Verdict counts over the 24 occupancy cells per metric and baseline (16 for the closed form: hands / head only)}} \\
\toprule
\headrow Metric & Baseline & better & not worse & n.s. & worse & insufficient \\
\midrule
AUROC & dwell table (label-free, train empirical) & 24 & 0 & 0 & 0 & 0 \\
AUROC & \cwname{} closed form (own hazards) & 16 & 0 & 0 & 0 & 0 \\
AUROC & persistence & 24 & 0 & 0 & 0 & 0 \\
Brier & dwell table (label-free, train empirical) & 18 & 4 & 2 & 0 & 0 \\
Brier & \cwname{} closed form (own hazards) & 16 & 0 & 0 & 0 & 0 \\
Brier & persistence & 24 & 0 & 0 & 0 & 0 \\
\bottomrule
\end{tabular*}\par\smallskip
\begin{tabular*}{\textwidth}{@{\extracolsep{\fill}}>{\tiny}l >{\tiny}l >{\tiny}l >{\tiny}l >{\tiny}r >{\tiny}l@{}}
\multicolumn{6}{@{}p{0.98\textwidth}@{}}{\textbf{(B) Engaged head-turned occupancy at 3\,s}} \\
\toprule
\headrow Metric & Baseline & future-state head (s0 / s1 / s2) & baseline (s0 / s1 / s2) & mean $\Delta$ $\pm$ sd & verdict \\
\midrule
Brier (lower better) & persistence & 0.0637 / 0.0635 / 0.0638 & 0.1009 / 0.1009 / 0.1009 & -0.0372$\pm$.0001 & better \\
Brier (lower better) & dwell table (label-free, train empirical) & 0.0637 / 0.0635 / 0.0638 & 0.0656 / 0.0656 / 0.0656 & -0.0019$\pm$.0001 & better \\
Brier (lower better) & \cwname{} closed form (own hazards) & 0.0637 / 0.0635 / 0.0638 & 0.0658 / 0.0661 / 0.0660 & -0.0023$\pm$.0003 & better \\
AUROC (higher better) & persistence & 0.7716 / 0.7771 / 0.7699 & 0.6602 / 0.6602 / 0.6602 & +0.1126$\pm$.0037 & better \\
AUROC (higher better) & dwell table (label-free, train empirical) & 0.7716 / 0.7771 / 0.7699 & 0.7406 / 0.7406 / 0.7406 & +0.0323$\pm$.0037 & better \\
AUROC (higher better) & \cwname{} closed form (own hazards) & 0.7716 / 0.7771 / 0.7699 & 0.7464 / 0.7429 / 0.7415 & +0.0293$\pm$.0045 & better \\
\bottomrule
\end{tabular*}\par\smallskip
\begin{tabular*}{\textwidth}{@{\extracolsep{\fill}}>{\tiny}l >{\tiny}l >{\tiny}l >{\tiny}r >{\tiny}r >{\tiny}r >{\tiny}l >{\tiny}r >{\tiny}l@{}}
\multicolumn{9}{@{}p{0.98\textwidth}@{}}{\textbf{(C) Recovery time and first-leave time, MAE in s (gaze off: own landmark chain, validation pending; no closed form)}} \\
\toprule
\headrow Regime & State & Anchors & $n$ & MAE head & MAE dwell table & verdict & MAE closed form & verdict \\
\midrule
engaged & hands on wheel & first leave (all anchors) & 449,350 & 0.3959$\pm$.0136 & 0.4907$\pm$.0000 & better & 0.4401$\pm$.0069 & better \\
engaged & hands on wheel & remaining dwell (off-task anchors) & 41,575 & 1.2013$\pm$.0401 & 1.3081$\pm$.0000 & n.s. & 1.3266$\pm$.0206 & better \\
engaged & head turned & first leave (all anchors) & 449,018 & 0.8168$\pm$.0067 & 0.9617$\pm$.0000 & better & 0.8526$\pm$.0117 & better \\
engaged & head turned & remaining dwell (off-task anchors) & 36,096 & 1.1660$\pm$.0203 & 1.3417$\pm$.0000 & better & 1.3178$\pm$.0196 & better \\
engaged & gaze off & first leave (all anchors) & 449,875 & 1.0509$\pm$.0192 & 1.2276$\pm$.0000 & better & -- & -- \\
engaged & gaze off & remaining dwell (off-task anchors) & 49,138 & 1.1520$\pm$.0159 & 1.2375$\pm$.0000 & better & -- & -- \\
manual & hands on wheel & first leave (all anchors) & 502,975 & 0.5439$\pm$.0218 & 0.5918$\pm$.0000 & n.s. & 0.5794$\pm$.0082 & better \\
manual & hands on wheel & remaining dwell (off-task anchors) & 34,770 & 1.3694$\pm$.0102 & 1.4324$\pm$.0000 & n.s. & 1.4138$\pm$.0175 & n.s. \\
manual & head turned & first leave (all anchors) & 503,483 & 0.8480$\pm$.0073 & 1.0090$\pm$.0000 & better & 0.8641$\pm$.0067 & better \\
manual & head turned & remaining dwell (off-task anchors) & 57,913 & 1.1679$\pm$.0215 & 1.3648$\pm$.0000 & better & 1.2688$\pm$.0101 & better \\
manual & gaze off & first leave (all anchors) & 500,166 & 1.0033$\pm$.0146 & 1.2453$\pm$.0000 & better & -- & -- \\
manual & gaze off & remaining dwell (off-task anchors) & 66,998 & 1.2110$\pm$.0071 & 1.3422$\pm$.0000 & better & -- & -- \\
\bottomrule
\end{cwtable*}

Xet Storage Details

Size:: 6.65 kB
Xet hash:: 3da73e498106e52519c0f9af8b388c3521c078a142610704cfb5fdcd1f3063ce

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.

	% tab_future_state.tex --- GENERATED by paper/tables/gen_tables.py --style iclr on 2026-09-18; do not edit by hand.
	% Every number is copied verbatim from these sources (path relative to the repository root : section / line):
	% - results/tables/future_state_eval.md : Verdict counts (table line 156)
	% - results/tables/future_state_eval.md : Occupancy BRIER, engaged head_turned 3s rows (table line 18)
	% - results/tables/future_state_eval.md : Occupancy AUROC, engaged head_turned 3s rows (table line 87)
	% - results/tables/future_state_eval.md : Recovery time / first-leave MAE (table line 209)
	% Full caption of the IEEE draft (provenance, verdict legend, line pointers; kept as a comment):
	% \| F2$'$ future-state head: a detached horizon-conditioned head ($\approx$0.034\,M params) on the FROZEN joint model's driver state $z_U$ + label-free
	% \| dwell features, trained on the train anchors of both regimes, selected on validation occupancy BCE; test split, seeds 0 / 1 / 2, device-cluster
	% \| bootstrap ($n_{\mathrm{boot}}{=}500$). States hands\_on / head\_turned (v2 causal kin targets) and gaze\_off (own landmark chain, validation pending).
	% \| Baselines: persistence $\pi_k = z_0$; dwell table = train empirical $P(s_{t+k} \mid z_0, \text{time-in-state bin})$; FINAL closed form = forward
	% \| marginal of the frozen model's own hazards (hands / head only). (A) Verdict counts over the cells regime $\times$ state $\times$ horizon (1 / 2 / 3 /
	% \| 5\,s); (B) engaged head\_turned at 3\,s; (C) MAE in s of $E[\min(T, 5.2\,\mathrm{s})]$ for the first-leave time (all anchors) and the remaining dwell
	% \| in the off-task state, paired $\Delta$ per seed and CIs in the source. $\Delta$ = head $-$ baseline; Brier / MAE lower better, AUROC higher. Verdicts
	% \| follow \texttt{kill\_checks.beats} ($\geq 2/3$ seeds CI excl.\ 0, same sign, $\|$mean $\Delta\| >$ seed sd): better / worse = significantly better /
	% \| worse; n.s.\ (INCONCLUSIVE) and not worse (NOT\_WORSE) = no significant difference. Nothing here changes the headline model. Sources:
	% \| future\_state\_eval.md verdict counts (line 156), Brier (line 18), AUROC (line 87), recovery time (line 209).
	\begin{cwtable*}{tab:future_state}{Detached future-state head on the frozen \cwname{} driver state (BATON test, 3 seeds): occupancy of hands-on / head-turned / gaze-off at 1--5\,s against persistence, a label-free dwell table and the model's own closed form. (A) Verdict counts over cells; (B) engaged head-turned at 3\,s; (C) recovery-time MAE in s. Better = significantly lower Brier or MAE, higher AUROC.\label{tab:tab_future_state}}{l l r r r r r}
	\multicolumn{7}{@{}p{0.98\textwidth}@{}}{\textbf{(A) Verdict counts over the 24 occupancy cells per metric and baseline (16 for the closed form: hands / head only)}} \\
	\toprule
	\headrow Metric & Baseline & better & not worse & n.s. & worse & insufficient \\
	\midrule
	AUROC & dwell table (label-free, train empirical) & 24 & 0 & 0 & 0 & 0 \\
	AUROC & \cwname{} closed form (own hazards) & 16 & 0 & 0 & 0 & 0 \\
	AUROC & persistence & 24 & 0 & 0 & 0 & 0 \\
	Brier & dwell table (label-free, train empirical) & 18 & 4 & 2 & 0 & 0 \\
	Brier & \cwname{} closed form (own hazards) & 16 & 0 & 0 & 0 & 0 \\
	Brier & persistence & 24 & 0 & 0 & 0 & 0 \\
	\bottomrule
	\end{tabular*}\par\smallskip
	\begin{tabular*}{\textwidth}{@{\extracolsep{\fill}}>{\tiny}l >{\tiny}l >{\tiny}l >{\tiny}l >{\tiny}r >{\tiny}l@{}}
	\multicolumn{6}{@{}p{0.98\textwidth}@{}}{\textbf{(B) Engaged head-turned occupancy at 3\,s}} \\
	\toprule
	\headrow Metric & Baseline & future-state head (s0 / s1 / s2) & baseline (s0 / s1 / s2) & mean $\Delta$ $\pm$ sd & verdict \\
	\midrule
	Brier (lower better) & persistence & 0.0637 / 0.0635 / 0.0638 & 0.1009 / 0.1009 / 0.1009 & -0.0372$\pm$.0001 & better \\
	Brier (lower better) & dwell table (label-free, train empirical) & 0.0637 / 0.0635 / 0.0638 & 0.0656 / 0.0656 / 0.0656 & -0.0019$\pm$.0001 & better \\
	Brier (lower better) & \cwname{} closed form (own hazards) & 0.0637 / 0.0635 / 0.0638 & 0.0658 / 0.0661 / 0.0660 & -0.0023$\pm$.0003 & better \\
	AUROC (higher better) & persistence & 0.7716 / 0.7771 / 0.7699 & 0.6602 / 0.6602 / 0.6602 & +0.1126$\pm$.0037 & better \\
	AUROC (higher better) & dwell table (label-free, train empirical) & 0.7716 / 0.7771 / 0.7699 & 0.7406 / 0.7406 / 0.7406 & +0.0323$\pm$.0037 & better \\
	AUROC (higher better) & \cwname{} closed form (own hazards) & 0.7716 / 0.7771 / 0.7699 & 0.7464 / 0.7429 / 0.7415 & +0.0293$\pm$.0045 & better \\
	\bottomrule
	\end{tabular*}\par\smallskip
	\begin{tabular*}{\textwidth}{@{\extracolsep{\fill}}>{\tiny}l >{\tiny}l >{\tiny}l >{\tiny}r >{\tiny}r >{\tiny}r >{\tiny}l >{\tiny}r >{\tiny}l@{}}
	\multicolumn{9}{@{}p{0.98\textwidth}@{}}{\textbf{(C) Recovery time and first-leave time, MAE in s (gaze off: own landmark chain, validation pending; no closed form)}} \\
	\toprule
	\headrow Regime & State & Anchors & $n$ & MAE head & MAE dwell table & verdict & MAE closed form & verdict \\
	\midrule
	engaged & hands on wheel & first leave (all anchors) & 449,350 & 0.3959$\pm$.0136 & 0.4907$\pm$.0000 & better & 0.4401$\pm$.0069 & better \\
	engaged & hands on wheel & remaining dwell (off-task anchors) & 41,575 & 1.2013$\pm$.0401 & 1.3081$\pm$.0000 & n.s. & 1.3266$\pm$.0206 & better \\
	engaged & head turned & first leave (all anchors) & 449,018 & 0.8168$\pm$.0067 & 0.9617$\pm$.0000 & better & 0.8526$\pm$.0117 & better \\
	engaged & head turned & remaining dwell (off-task anchors) & 36,096 & 1.1660$\pm$.0203 & 1.3417$\pm$.0000 & better & 1.3178$\pm$.0196 & better \\
	engaged & gaze off & first leave (all anchors) & 449,875 & 1.0509$\pm$.0192 & 1.2276$\pm$.0000 & better & -- & -- \\
	engaged & gaze off & remaining dwell (off-task anchors) & 49,138 & 1.1520$\pm$.0159 & 1.2375$\pm$.0000 & better & -- & -- \\
	manual & hands on wheel & first leave (all anchors) & 502,975 & 0.5439$\pm$.0218 & 0.5918$\pm$.0000 & n.s. & 0.5794$\pm$.0082 & better \\
	manual & hands on wheel & remaining dwell (off-task anchors) & 34,770 & 1.3694$\pm$.0102 & 1.4324$\pm$.0000 & n.s. & 1.4138$\pm$.0175 & n.s. \\
	manual & head turned & first leave (all anchors) & 503,483 & 0.8480$\pm$.0073 & 1.0090$\pm$.0000 & better & 0.8641$\pm$.0067 & better \\
	manual & head turned & remaining dwell (off-task anchors) & 57,913 & 1.1679$\pm$.0215 & 1.3648$\pm$.0000 & better & 1.2688$\pm$.0101 & better \\
	manual & gaze off & first leave (all anchors) & 500,166 & 1.0033$\pm$.0146 & 1.2453$\pm$.0000 & better & -- & -- \\
	manual & gaze off & remaining dwell (off-task anchors) & 66,998 & 1.2110$\pm$.0071 & 1.3422$\pm$.0000 & better & -- & -- \\
	\bottomrule
	\end{cwtable*}