The Decision Value of Perception Compute

Hoang Pham Cong; Ho Viet Duc Luong

cs.LGSep 28, 2026

The Decision Value of Perception Compute

Authors: Hoang Pham Cong, Ho Viet Duc Luong

Organizations: Hanoi University of Science and Technology (HUST), Hanoi, Vietnam

Abstract

Adaptive perception spends extra computation on inputs where perception is expected to improve. When perception feeds a downstream decision system, a better perception output need not produce a better decision. We define the decision value of perception compute as the change in downstream loss from escalating an input from a cheap to an expensive perception mode. Because this value can be negative, the allocation of perception compute should be judged against a budget-constrained decision oracle, with uniform full-fidelity inference as a baseline rather than an upper bound. We introduce DEEP (Decision Evaluation for Escalated Perception), a benchmark that scores pre-escalation allocators against this oracle under selection, latency and energy budgets, charging each allocator for its own computation. With deployed monocular geometry on KITTI and nuScenes, we find that 34--54% of the escalations that change downstream loss make it worse; harmful escalations also occur for the published PDM-Closed planner, evaluated open-loop on nuPlan with real detector outcomes. On nuScenes, perception-level gain frequently disagrees in sign with decision value. This mismatch has practical consequences: choosing among fixed deployable signals by missed-object perception gain rather than by decision value reduces realized test decision gain by 7.4% of the all-cheap loss on average. Learned allocators recover part of the oracle's value by finding beneficial escalations but select nearly as much harm as random, and once their own computation is charged at a 20% latency budget, only the lightweight routers, at about 3.5% of a full detector pass, still beat random.

Figures & tables

Panel of Figure 1 — Figure 1: Better detection need not improve the decision. Two nuScenes frames under the same braking controller and monocular geometry. In both, the full mode recovers a bus that the cheap mode misses (green box; dashed yellow: the missed reference object); the shaded band is the controller’s driving corridor on the road. (a) The recovered bus lies outside the corridor, but the monocular lift places it inside, and the controller brakes hard. (b) The recovered bus lies inside the corridor, and the controller decelerates. Both improve detection, yet their decision values have opposite signs.

Figure 2: The DEEP task. Top, deployment: every input pays for the cheap mode and the allocator, which scores it from pre-escalation information only; an escalated input also pays for its higher-fidelity mode. Bottom, offline evaluation, unavailable to the allocator: the decisions obtained from the cached cheap and high-fidelity outputs are scored on the same reference scene state to give $V_{i}^{q}$ , inputs are selected from the scores under the budget, and $\mathrm{nDG}=$ realized gain / oracle gain.

Dataset, system, geometry	Transition / loss	Affected	Harmed	$\rho_{q}$	All-full	Oracle@20
nuScenes, $q_{\mathrm{brake}}$ , oracle	YOLOv8s 320 $\to$ 640	141	35.5%	0.42	14.01%	24.07%
nuScenes, $q_{\mathrm{brake}}$ , mono	YOLOv8s 320 $\to$ 640	273	39.6%	0.41	16.47%	27.77%
nuScenes, $q_{\mathrm{plan}}$ , oracle †	YOLOv8s 320 $\to$ 640	1,179	50.9%	0.70	0 1.31%	0 4.41%
nuScenes, $q_{\mathrm{plan}}$ , mono †	YOLOv8s 320 $\to$ 640	1,612	49.3%	0.71	0 1.59%	0 5.31%
KITTI, $q_{\mathrm{traj}}$ , mono	YOLOv8s 320 $\to$ 640	678	48.4%	0.28	12.78%	17.81%
KITTI, $q_{\mathrm{traj}}$ , mono	YOLOv8s 384 $\to$ 640	458	48.7%	0.46	0 5.97%	11.03%

Table 1: Sign-varying decision value across datasets, detector families, fidelity gaps and downstream systems, over all units. Affected: inputs with non-zero decision value, of 3,376 nuScenes frames (2,655 for

q_{\mathrm{plan}}

), 8,008 KITTI frames and 1,440 nuPlan states. Harmed: share of affected inputs made worse by the full mode. All-full and Oracle@20: loss reduction relative to all-cheap inference, for escalating every input (all-full, a compute-agnostic baseline) and for the oracle escalating up to 20% of inputs; † marks budget-binding cells. Rows show both geometries on nuScenes and deployed geometry on KITTI; the KITTI oracle-geometry controls are in Appendix C . nuPlan rows use real detector outcomes on nuPlan images. See also Appendices C and F .

	Controls		Heuristic		Gate		Detection-list router
Cell	Rand.	Ego sp.	Unc.	Crit.	ridge	GBM	MLP, $V$	MLP, $V{>}0$	GBM, $V$	GBM, $V{>}0$
nuScenes oracle, $q_{\mathrm{brake}}$	0.148	0.208	0.015	0.470	0.254	0.155	0.354	0.354	0.196	0.421
nuScenes mono, $q_{\mathrm{brake}}$	0.132	0.148	0.088	0.407	0.293	0.482	0.366	0.255	0.383	0.369
nuScenes oracle, $q_{\mathrm{plan}}$	$-0.006$	$-0.021$	$-0.117$	$-0.272^{*}$	$-0.150$	$-0.141$	$-0.079$	$-0.074$	$-0.083$	0.031
nuScenes mono, $q_{\mathrm{plan}}$	0.092	$-0.015$	0.013	$-0.060$	0.063	0.042	0.250	0.102	$-0.022^{*}$	0.090
KITTI oracle, $q_{\mathrm{brake}}$	0.139	0.584	$-0.010^{*}$	$-0.013^{*}$	0.238	0.226	0.255	0.313	0.293	0.360

Table 2: Official held-out results on the core track (YOLOv8s 320

\to

640): nDG of the deployable baselines at a 20% selection budget on the frozen test splits. Bold (significance, not rank): the paired 95% interval of realized decision gain against random, over every bootstrap draw, lies above zero; ∗ : below zero. Planner rows use

q_{\mathrm{plan}}

with average displacement; the final-displacement variants, which with these rows form the ten core cells, are in Appendix E . The detection-list router appears in its four pre-specified variants, fitted to

V

\mathbf{1}[V>0]

Figure 3: Where decision value is lost , in shares of the all-cheap loss unless stated. (a) Benefit captured (green), harm selected (red) and net gain (black) of each allocator at a 20% selection budget, mean over the ten core cells. (b) nDG with the allocator free and with its measured latency charged, 20% budget, mean over the ten core cells; inf.: the allocator’s cost exceeds the budget. (c) Test gain of the signal chosen on validation by decision value minus that of the signal chosen by false-negative gain, among the target-free signals of Section 5.4 ; markers: the decision objective’s choice (open: same choice); paired 95% intervals, red if excluding zero; pooled: their equal-weight mean.

Appendix figures & tables23 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 4 — Figure 4: Perception-level signals recover part of the decision value, and the best one depends on the downstream system. nDG on nuScenes (YOLOv8s 320 $\to$ 640, oracle geometry, all units) when escalation is ranked by each signal, computed exactly after escalation. Right: the planner scored against its own reference-conditioned output.

		Braking	Published learned planner
Allocation score	class	$q_{\mathrm{brake}}$	$q_{\mathrm{plan}}$	$q_{\mathrm{plan}}^{\,\mathrm{self}}$
Random (16 seeds)	deployable	0.122	0.061	0.123
Cheap-detection uncertainty	deployable	0.005	0.134	0.096
Pre-escalation gate, GBM	deployable	0.309	0.079	—
Criticality (reference geometry)	diagnostic	0.182	0.016	0.219
False-negative gain $G^{\mathcal{E}_{\mathrm{FN}}}$	diagnostic	0.171	0.228	0.328

Appendix

Table 3: Cross-target diagnostic analysis: nDG at a 20% budget on nuScenes, oracle geometry, all units. Only the first three rows are deployable; the gate is fit out of fold by scene. The

q_{\mathrm{plan}}^{\,\mathrm{self}}

column uses all 3,376 frames, the

q_{\mathrm{plan}}

column the 2,655 frames with a logged future; the self-reference gap in the text recomputes

q_{\mathrm{plan}}^{\,\mathrm{self}}

on those 2,655 frames. Bold: best diagnostic score in each column.

Downstream system	Pair	Geometry	Affected	Harmed	$\rho_{q}$	All-full / Oracle@20
nuScenes, $q_{\mathrm{brake}}$	Y8 320 $\to$ 640	oracle	141	35.5%	0.42	14.01 / 24.07%
nuScenes, $q_{\mathrm{plan}}$	Y8 320 $\to$ 640	oracle	1,179	50.9%	0.70	0 1.31 / 0 4.41%
nuScenes, $q_{\mathrm{plan}}^{\,\mathrm{self}}$	Y8 320 $\to$ 640	oracle	1,413	40.9%	0.46	0 7.49 / 13.67%
KITTI, $q_{\mathrm{traj}}$	Y8 320 $\to$ 640	mono	0, 678	48.4%	0.28	12.78 / 17.81%
KITTI, $q_{\mathrm{traj}}$	Y8 320 $\to$ 640	oracle	0, 463	0 3.0%	0.003	68.34 / 68.58%
KITTI, $q_{\mathrm{traj}}$	Y8 384 $\to$ 640	mono	0, 458	48.7%	0.46	0 5.97 / 11.03%

Appendix

Table 4: Sign-varying decision value and harm ratio across detector families, fidelity gaps and geometries. Loss reductions are relative to all-cheap inference; Oracle@20 has capacity up to 20% of inputs. Every KITTI row is budget-slack, so

\rho_{q}=1-{}

all-full

/

Oracle@20 there.

\rho_{q}

is smallest for the largest fidelity gap under oracle geometry (row 5), where escalation is almost always beneficial.

	$q_{\mathrm{brake}}$				$q_{\mathrm{traj}}$
	Harmed		$\rho_{q}$		Harmed		$\rho_{q}$
Pair	mono	oracle	mono	oracle	mono	oracle	mono	oracle
YOLOv8s 320 $\to$ 640	34.2	24.6	0.168	0.080	48.4	3.0	0.282	0.003
YOLOv8s 384 $\to$ 640	39.8	35.3	0.473	0.295	48.7	5.1	0.459	0.002
YOLOv8s 512 $\to$ 640	42.9	39.7	0.615	0.467	53.8	17.0	1.114	0.051
RT-DETR-l 320 $\to$ 640	44.2	47.5	0.701	0.691	49.7	19.9	0.762	0.023

Appendix

Table 5: Harm with deployed monocular and oracle geometry on every KITTI fidelity pair, over all units. Harmed: share of affected inputs made worse by the full mode (%);

\rho_{q}

: harm ratio.

Cell	S0 (shared)	S1 (equal boxes)	S2 (F1)	S3 (precision)
nuScenes oracle, $q_{\mathrm{brake}}$	35.5% / 0.42	38.5% / 0.54	34.5% / 0.39	34.1% / 0.41
nuScenes oracle, $q_{\mathrm{plan}}$	50.9% / 0.70	50.2% / 0.78	50.1% / 0.71	50.2% / 0.72
nuScenes mono, $q_{\mathrm{brake}}$	39.6% / 0.41	36.2% / 0.34	37.0% / 0.34	37.5% / 0.37
nuScenes mono, $q_{\mathrm{plan}}$	49.3% / 0.71	49.6% / 0.83	49.3% / 0.73	49.2% / 0.71
KITTI Y8 384 $\to$ 640, $q_{\mathrm{traj}}$	48.7% / 0.46	48.1% / 0.48	48.2% / 0.47	48.6% / 0.47
KITTI Y8 384 $\to$ 640, $q_{\mathrm{brake}}$	39.8% / 0.47	34.9% / 0.44	37.1% / 0.45	38.8% / 0.47

Appendix

Table 6: Harm rate and harm ratio

\rho_{q}

with a shared threshold (S0) and with per-mode operating points (S1–S3), over all units. KITTI rows use monocular geometry.

	Detection-list router
Cell	MLP, $V$	MLP, $V{>}0$	GBM, $V$	GBM, $V{>}0$	Pixel router
nuScenes oracle, $q_{\mathrm{brake}}$	0.354	0.354	0.196	0.421	$-0.048$
nuScenes oracle, planner ADE	$-0.079$	$-0.074$	$-0.083$	0.031	$-0.163$
nuScenes oracle, planner FDE	0.037	$-0.047$	$-0.093$	0.065	$-0.034$
nuScenes mono, $q_{\mathrm{brake}}$	0.366	0.255	0.383	0.369	0.215
nuScenes mono, planner ADE	0.250	0.102	$-0.022^{*}$	0.090	0.082

Appendix

Table 7: Router variants, nDG at a 20% budget on the held-out test cells. Bold: the paired 95% interval of realized decision gain against random, over every bootstrap draw, lies above zero; asterisk: below zero. nuPlan rows use real detector outcomes; the pixel router does not apply to nuPlan.

Cell	Random	Ego speed	Uncertainty	Criticality	Gate, ridge	Gate, GBM
nuScenes oracle, $q_{\mathrm{brake}}$	0.148	0.208	0.015	0.470	0.254	0.155
nuScenes oracle, planner ADE	$-0.006$	$-0.021$	$-0.117$	$-0.272^{*}$	$-0.150$	$-0.141$
nuScenes oracle, planner FDE	0.059	0.024	$-0.016$	$-0.142$	0.030	0.125
nuScenes mono, $q_{\mathrm{brake}}$	0.132	0.148	0.088	0.407	0.293	0.482
nuScenes mono, planner ADE	0.092	$-0.015$	0.013	$-0.060$	0.063	0.042
nuScenes mono, planner FDE	0.080	0.007	$-0.059$	$-0.048$	$-0.043^{*}$	0.079

Appendix

Table 8: Deployable baselines, nDG at a 20% budget on all held-out test cells. Bold: the paired 95% interval of realized decision gain against random, over every bootstrap draw, lies above zero; asterisk: below zero. nuPlan rows use the real-perception track; † no uncertainty signal is defined on it. Planner rows are

q_{\mathrm{plan}}

with average (ADE) and final (FDE) displacement.

Cell	Criticality (reference)	$G^{\mathcal{E}_{\mathrm{FN}}}$	$G^{\mathcal{E}_{\mathrm{risk}}}$	PKL	TIP
nuScenes oracle, $q_{\mathrm{brake}}$	0.287	0.258	0.687	0.461	0.365
nuScenes oracle, planner ADE	$-0.195$	0.123	$-0.085$	$-0.132$	$-0.181$
nuScenes oracle, planner FDE	$-0.005$	0.121	$-0.036$	0.124	$-0.035$
nuScenes mono, $q_{\mathrm{brake}}$	0.323	0.210	0.330	0.215	0.110
nuScenes mono, planner ADE	$-0.051^{*}$	0.350	0.404	0.126	0.117
nuScenes mono, planner FDE	$-0.037$	0.264	0.246	0.077	0.168

Appendix

Table 9: Diagnostic signals, nDG at a 20% budget on the held-out test cells. Each column is one fixed signal, not a per-cell selection. Bold: the paired 95% interval of realized decision gain against random, over every bootstrap draw, lies above zero; asterisk: below zero. These signals require the full output or the reference state. PKL and TIP are defined only on nuScenes. On the real-perception nuPlan track the false-negative gain counts missed tracks; neither it nor the risk-weighted gain sees the false positives that drive most of PDM-Closed’s decision value.

Panel of Figure 5 — Figure 5: Charging allocator cost reorders the allocators. nDG on the held-out split at a 20% budget, when the allocator is free (open) and when its measured latency on the device is charged (filled; the gradient-boosted gate with batched inference). The feature-based gates lose their advantage because feature extraction consumes the budget (inf.: the allocator cannot run within it at all), while both MLP variants of the detection-list router, at about 3.5% of a full pass, keep most of it; its gradient-boosted variants cannot run within any measured budget. Figure 6 covers every budget level.

Panel of Figure 6 — Figure 6: nDG across measured latency budgets. Held-out split, the three core cells of Figure 5 , every budget level. Missing points are infeasible: the ridge gate cannot run within the 10 and 20% budgets, the batched gradient-boosted gate within 10%, and the pixel router within 10 and 20% on KITTI and any budget on nuScenes. Both MLP variants of the detection-list router, at about 3.5% of a full pass, keep their ranking skill.

Cell	Half-width	Cell	Half-width
nuScenes oracle, planner ADE	0.7%	KITTI mono, $q_{\mathrm{brake}}$	3.7%
nuPlan IDM, scalar	1.0%	nuScenes mono, $q_{\mathrm{brake}}$	5.8%
nuScenes oracle, planner FDE	1.1%	nuScenes oracle, $q_{\mathrm{brake}}$	6.0%
nuScenes mono, planner ADE	1.1%	KITTI oracle, $q_{\mathrm{brake}}$	6.5%
nuPlan IDM, safety	1.1%	nuPlan PDM-Closed, scalar	14.0%
nuScenes mono, planner FDE	1.7%	nuPlan PDM-Closed, safety	16.0%

Appendix

Table 10: Typical uncertainty of paired comparisons with random at a 20% budget: the median half-width, over all scored signals, of the paired 95% interval on realized decision value, as a share of the all-cheap loss, with every bootstrap draw kept.

Figure 7: Target transform against training objective. Difference in nDG from the same regression router trained on raw decision value, on the ten core cells at a 20% budget; bold with an asterisk: paired 95% interval excludes zero. Rank and signed CDF keep the decision-value objective and change only its transform; ORIC’s reward and the AP-swap objective change the objective in its published CDF form. Post hoc, uncorrected for multiplicity.

Allocator	Overhead per input	Escalated at 20% budget	at 50% budget
Random	0 ms	20.0 / 20.0%	50.0 / 50.0%
Cheap-detection uncertainty	0.21–0.29 ms	18.5 / 18.8%	48.5 / 48.8%
Cheap-side criticality	1.30–1.47 ms	12.4 / 12.9%	42.4 / 42.9%
Detection-list router, MLP	0.65 ms	16.6 / 16.5%	46.6 / 46.5%
Gate, ridge	3.93 ms	inf. / inf.	29.7 / 28.7%
Gate, GBM, batched inference	3.54 ms	1.7 / 0.8%	31.7 / 30.8%

Appendix

Table 11: Share of inputs each deployable allocator can escalate under a measured per-input latency budget after paying for itself, on nuScenes / KITTI (median over cells). inf.: the allocator’s own overhead exceeds the budget,

C_{S}>b-C_{\mathsf{c}}

, so it cannot run at all. Diagnostic signals require the full output or the reference state and are not deployable under measured budgets.

Planner	Loss	Affected	Harmed	$\rho_{q}$	All-full / Oracle@20
Real detector outcomes
PDM-Closed	collision	24	25.0%	0.33	11.54 / 17.31%
PDM-Closed	clearance shortfall	45	44.4%	0.34	0 5.22 / 0 7.89%
PDM-Closed	logged-trajectory deviation	104	37.5%	0.55	0 1.90 / 0 4.25%
PDM-Closed	safety aggregate	45	44.4%	0.33	0 9.81 / 14.74%
PDM-Closed	scalar aggregate	104	37.5%	0.35	0 8.26 / 12.68%

Appendix

Table 12: All five losses for both published planners under real detector outcomes, over all 1,440 states, and the two aggregate losses under the KITTI detection profile. The losses are geometric and are not nuPlan’s metric suite. The IDM collision row is a single affected state.

	nuScenes, $q_{\mathrm{brake}}$			nuScenes, $q_{\mathrm{plan}}^{\,\mathrm{self}}$
	helped	harmed	unaffected	helped	harmed	unaffected
$\Delta$ false positives	$+1.176$	$+1.980$	$+0.849$	$+1.093$	$+1.251$	$+0.671$
$\Delta$ false negatives	$-1.396$	$-0.600$	$-0.790$	$-1.321$	$-1.014$	$-0.522$
$\Delta$ detections	$+2.505$	$+2.480$	$+1.581$	$+2.315$	$+2.190$	$+1.154$
frames with an added FP	61.5%	78.0%	—	60.0%	64.2%	—
frames recovering no miss	25.3%	56.0%	—	31.0%	44.6%	—

Appendix

Table 13: Mean change in detection errors from cheap to full mode (

\Delta=

full

-

cheap), by the sign of decision value. Harmed frames add more false positives and recover fewer misses.

Figure 8: Where escalation changes what the braking controller sees. Bird’s-eye view of every object the controller considers — inside its corridor (shaded) and within 80 m — that differs between the cheap and full modes, on the nuScenes frames with non-zero decision value under oracle geometry whose corridor objects differ between the modes. Harmed frames are dominated by false positives that the full mode adds; helped frames by misses that it recovers.

Figure 9: Perception gain does not determine decision value. Each point is a nuScenes frame with non-zero perception gain and decision value (YOLOv8s 320

\to

640, oracle geometry), on

\operatorname{asinh}

axes. Shaded quadrants mark sign disagreement: 32.6% of frames for the braking controller and 50.6% for the learned planner.

	$q_{\mathrm{brake}}$ (141 frames)			$q_{\mathrm{plan}}$ (1,179 frames)
Perception gain	Corr.	$P(\mathrm{harm}{\mid}G{>}0)$	Disagree	Corr.	$P(\mathrm{harm}{\mid}G{>}0)$	Disagree
Exact (false-negative count)	$+0.08$	24.4%	25.0%	$+0.02$	50.6%	50.4%
False negatives + false positives	$+0.07$	15.7%	31.9%	$+0.02$	50.3%	49.8%
Combined error ( $\mathcal{E}_{5}$ )	$+0.08$	19.2%	32.6%	$+0.01$	51.0%	50.6%
Risk-weighted ( $\mathcal{E}_{\mathrm{risk}}$ )	$+0.11$	23.8%	26.8%	$+0.01$	50.5%	51.0%

Appendix

Table 14: Perception-level gains do not fix the sign of decision value. nuScenes, YOLOv8s 320

\to

640, oracle geometry, shared threshold, frames with non-zero decision value. Corr.: Spearman correlation between gain and decision value.

P(\mathrm{harm}{\mid}G{>}0)

: share of frames with positive perception gain that escalation harms. Disagree: share of frames with non-zero gain and non-zero decision value whose signs differ. Per-mode thresholds are in Appendix D .

Panel of Figure 10 — Figure 10: Escalation adds detections the controller reacts to, and recovers the objects it needs. Braking controller on nuScenes, monocular geometry; each pair is cropped around the object that decides the braking action. Left pairs: harmed frames in which the full mode adds a detection with no reference object (red) that triggers hard braking. Right pairs: helped frames in which the full mode detects an object (green) that the cheap mode misses (dashed yellow). Thin white boxes are the other detections. More examples and the selection rule are in Figure 11 .

Panel of Figure 11 — Figure 11: Qualitative examples under the braking controller (nuScenes, monocular geometry), extending Figure 10 ; colors as there, each pair cropped around the object that decides the braking action. The examples are the frames among the six most harmed and six most helped, at most one per scene, whose decisive change is an added detection or a recovered object; in the other four, escalation changes the estimated range of an object that both modes detect. The first harmed frame is harmed by a recovered object: the full mode detects a vehicle that the cheap mode misses, and the monocular lift places it inside the controller’s corridor although its reference box lies outside it. In the second helped frame, the full mode’s box around the bus overlaps the reference box by less than the matching threshold and is therefore drawn as an added detection.

Configuration	Task	nDG of $G^{\mathcal{E}_{\mathrm{FN}}}$	Multi-metric diagnostic
KITTI/Y8 320 $\to$ 640 mono	longitudinal	0.162	0.672
KITTI/Y8 320 $\to$ 640 mono	lateral	$-0.230$	$-0.118$
KITTI/Y8 320 $\to$ 640 oracle	longitudinal	0.200	0.810
KITTI/Y8 320 $\to$ 640 oracle	lateral	$-1.041$	$-0.231$
KITTI/Y8 384 $\to$ 640 mono	longitudinal	0.152	0.388
KITTI/Y8 384 $\to$ 640 mono	lateral	$-0.297$	$-0.273$

Appendix

Table 15: nDG of the false-negative gain (the reduction in missed objects) and of a multi-metric diagnostic at a 20% budget, across 16 configurations. The diagnostic is an out-of-fold gradient-boosted model over perception-error features computed from the reference state, so it is not an upper bound over perception signals: in four configurations it falls below the false-negative gain beside it. It exceeds

0.5

only in the two aggressive-gap KITTI longitudinal cells and in nuScenes longitudinal under oracle geometry.

	Braking weights		Braking corridor		Trajectory weights
Setting	Harmed	$\rho_{q}$	Harmed	$\rho_{q}$	Harmed	$\rho_{q}$
nuScenes, oracle	35.5–36.2	0.28–0.62	35.5–36.5	0.33–0.46	—
nuScenes, mono	38.1–42.5	0.28–0.55	39.2–41.6	0.37–0.41	—
KITTI Y8 320 $\to$ 640, mono	32.6–35.3	0.11–0.28	32.7–34.6	0.15–0.22	46.0–48.8	0.27–0.32
KITTI Y8 384 $\to$ 640, mono	38.8–40.1	0.35–0.65	38.9–39.8	0.41–0.51	47.3–51.0	0.45–0.51
KITTI Y8 512 $\to$ 640, mono	42.9–43.4	0.51–0.72	42.6–42.9	0.59–0.71	51.6–54.5	1.10–1.15

Appendix

Table 16: Harm rate (%) and harm ratio

\rho_{q}

across loss and corridor variants, as the range over variants that include the default, over all units. Braking weights: over-braking, collision and change-of-command weights; corridor: half-width 1.0–1.5 m; trajectory weights: collision, clearance and progress weights in planning and loss.

Mode	GPU ms	End-to-end ms	Energy mJ
YOLOv8s 320	5.09	13.18	41.1
YOLOv8s 384	6.09	14.38	47.8
YOLOv8s 512	7.89	16.72	62.0
YOLOv8s 640	9.28	18.47	84.4
RT-DETR-l 320	16.82	25.31	397
RT-DETR-l 480	21.35	31.21	535

Appendix

Table 17: Jetson AGX Xavier FP16 TensorRT profiling of the KITTI detector modes on an idle board, one image per call. End-to-end time includes image decoding and preprocessing. Energy is per frame under the module convention, the GPU, CPU and SoC rails over idle.

Explore similar work

Jul 31, 2026cs.RO

Belief-Space Perception Routing under Coupled Sensor Faults and Compute Contention

A robot that has to see and react on a fixed clock runs into two problems at once. Its cameras degrade in rain, mud, fog, and darkness. And the single onboard processor it runs on is shared with planning and control, so the compute left over for perception moves around from second to second. Most systems model the two separately. We present a perception router that tracks probabilistic estimates of sensor-fault state and compute- contention state, couples them with a noisy-OR term, and uses the coupled estimate to pick one of four detector configurations (YOLO11x/n at 1280 or 640 px) so that the frame finishes before its deadline. Where the two stressors co-occur, the coupled policy cuts the deadline-miss rate by 1.1 to 9.4 percentage points against a policy that treats them independently. The interval excludes zero in five of six conditions, the pooled effect over 10 sequences and 6 conditions has sign-test p = 0.001, and every uncoupled control and the fault-free trajectory sit at exactly 0.0 pp. Routing costs tens of microseconds per frame. We then asked whether the coupling the method exploits arises on its own. Across eight real RADIATE adverse-weather sequences and three workload proxies independent of the fault signal, after Benjamini-Hochberg correction and a replication run, none of 24 tests found it. We report that null and scope the routing result as a proof of mechanism. Whether such coupling occurs in the field is still open, and the released evaluation pipeline lets a deployment settle it on its own traces.

Sparsh Roy, Vihan Aggarwal, Davin Yin

Massachusetts Institute of Technology, Cambridge, MA, USA · Hopewell Valley Central High School, New Jersey, USA · Millburn High School, New Jersey, USA +1

Collaborative Perception Robotic Perception Perception Fault Diagnosis

Ask PDF

Jun 29, 2026cs.NI

Budget-Adaptive Routing: Skipping the Weak When the Strong Answers Anyway

Edge-cloud inference collaborations are often designed with a routing estimator that decides whether to offload each frame from weak models at the edge to stronger models in the cloud. Existing systems place the routing estimator after the weak detector, so the weak forward pass still runs even on frames that are later offloaded. In this paper, we argue that this weak-conditioned design can be suboptimal when the offload budget varies. First, we present a competitive weak-skipping estimator (0.153 GFLOPs, about 29x lighter than the weak detector at 4.49 GFLOPs) that extracts routing signal from raw pixels, outperforming the common after-weak placement weak-conditioned baselines. Second, we show that neither weak-skipping nor weak-conditioned placement dominates across the full operating curve, and we propose budget-adaptive routing, which selects between them by offload budget via two offline-tuned thresholds. On PASCAL VOC, our budget-adaptive router traces the upper accuracy envelope of both fixed placements across the operating range. Our method reduces per-frame latency by up to 19.1 ms (about 30% lower at rho = 0.9). Besides outperforming SOTA methods, it is surprisingly stronger than the strong model (+1.7 pp over the strong model's peak mAP) at some operating points with far less compute. Artifacts are available at https://github.com/ViGeng/bgt-ada

Wei Geng, Nitinder Mohan, Jörg Ott

Technical University of Munich Munich, Germany · TU Delft Delft, Netherlands

Answer

Ask PDF

Jul 30, 2026cs.AI

Search as Computation Allocation

Many algorithms spend an internal resource before returning a decision and are evaluated only by the quality of that terminal output. We formalize such procedures as terminal computation-allocation problems: costly computations produce observations, update beliefs about a latent environment, and matter only through terminal decision loss. Bellman equations characterize optimal allocation under fixed budgets, priced computation, and exact certification. We then relate value of computation (VOC) to information. Mutual information equals myopic VOC under log loss, whereas under simple regret VOC is a knowledge-gradient quantity; moreover, information gain can rank computations arbitrarily poorly, although it gives a one-sided upper bound on VOC. Bandit pulls, tree simulations, and node expansions illustrate the same model under different computation topologies. Finally, under an explicit frontier-resolution and heuristic-error model, maximizing approximate VOC recovers weighted A*, with A* and greedy best-first search as limiting cases. The theory identifies a shared decision problem without asserting that one acquisition rule is universally optimal.

Alexander Tuisov

Technion - Israel Institute of Technology

Computational Cost Bellman Equation

Ask PDF