KEPT: Conformal Keepouts Kept in Planning for Enclosed-Space Navigation in Degraded Sensing

Eric Minwoo Kim; Jong-Kook Kim

cs.ROOct 27, 2026

KEPT: Conformal Keepouts Kept in Planning for Enclosed-Space Navigation in Degraded Sensing

Authors: Eric Minwoo Kim, Jong-Kook Kim

Organizations: School of Electrical Engineering, Korea University, Seoul, Republic of Korea

Abstract

Autonomous vehicles in enclosed spaces such as tunnels and mines rely on range maps that smoke, dust and fog make inaccurate. Conformal prediction calibrates a keepout around the mapped walls that covers the true unsafe set at a chosen probability. Typically an unmodified planner treats the keepout as an inflated obstacle, and nothing ensures that the vehicle can stop before the keepout as it moves when sensing degrades. In narrow passages this margin and the planner's buffer close the route, and existing remedies recover reach by relaxing the coverage level. This paper conditions the conformal margin on a label-free severity estimate to hold coverage near its target as degradation varies. A planning and control layer enforces one braking-distance condition on the keepout at plan time and at every command, takes the measured tracking residual as its only margin and never alters the keepout. The keepout is validated on ground-robot smoke recordings and the closed loop in simulated quadrotor flights in a foggy tunnel against the standard deployment. In closed-loop simulation the proposed system substantially reduces entries into the calibrated keepout and reaches numerically more goals than the deployment pattern in both calibration regimes, an observed gain of 3.2 percentage points pooled whose per-regime confidence intervals include zero. The per-class keepout holds its coverage near the target on held-out recordings.

Figures & tables

Symbol	Meaning
$W$ , $O$ , $\varphi(x)$	workspace, true obstacle set, true distance to $O$
$r_{\mathrm{safe}}$ , $U$	vehicle radius, true unsafe set of centre positions
$\Delta$ , $\mathrm{cell}(x)$	grid edge, cell that contains $x$
$m(x)$	map estimate of the distance to the nearest surface
$\hat{\beta}(\theta)$ , $b(x)$	severity field per azimuth wedge, stratum of a cell
$z$ , $\alpha$ , $n_{b}$ , $\lambda_{b}$	score, target miscoverage, calibration count and margin of stratum $b$

TABLE I: Notation of Sections III and IV .

Arm	Definition	Role
M	plan-time enforcement and command filter, the planning layer of KEPT (Section IV )	the method
L	the plan-time law alone, no command filter	what plan-time enforcement adds
F	the command filter alone, in front of the $A^{\prime}$ planner	what the command filter adds
$A^{\prime}$ 0.15	the stock planner on the certified map, soft clearance 0.15 m and its own emergency stop	the conformal deployment pattern, comparator of the claim
$A^{\prime}$ 0.25	the same at soft clearance 0.25 m	a second point on the clearance curve of $A^{\prime}$
$A^{\prime}$ 1.10	$A^{\prime}$ 0.15 at a speed cap of 1.10 m/s	whether the result is an effect of speed

TABLE II: Arms of the closed-loop evaluation. All arms share the certificate, the poses, the seeds, the goal and the planner limits.

Regime	Stratum	$n$	EF	EP	R	$\Delta$ EF [95 %] $p$	$\Delta$ R [95 %] $p$	Won
Simulator-calibrated	clear 0.500	24	1 / 1	2 / 6	13 / 21	$+0.00$ [-0.25, +0.25] 1.00	$-0.67$ [-1.33, +0.08] 0.16	no
	clear 0.400	24	2 / 1	6 / 1	19 / 13	$+0.08$ [-0.17, +0.33] 1.00	$+0.50$ [-0.08, +1.08] 0.22	no
	fog 0.500	21	0 / 3	0 / 5	14 / 14	$-0.25$ [-0.50, +0.00] 0.25	$+0.00$ [-0.42, +0.42] 1.00	no
	fog 0.400	24	0 / 4	0 / 4	19 / 13	$-0.33$ [-0.58, -0.08] 0.12	$+0.50$ [+0.08, +0.92] 0.11	yes
	pooled	93	3 / 9	8 / 16	65 / 61	$-0.13$ [-0.25, +0.00] 0.15	$+0.08$ [-0.19, +0.38] 0.70	yes
Transferred	clear 0.500	24	2 / 3	2 / 7	15 / 20	$-0.08$ [-0.25, +0.00] 1.00	$-0.42$ [-0.92, +0.08] 0.25	no

TABLE III: The proposed stack (M) against the deployment pattern (

A^{\prime}

0.15) on the matched cells of each regime. EF entry flights, EP entry episodes, R goals reached, M /

A^{\prime}

\Delta

EF and

\Delta

R are the paired per-pose differences M minus

A^{\prime}

, their 95 % bootstrap intervals and the permutation

p

. Won marks the pre-registered two-requirement criterion, fewer entries and at least as many goals.

Regime	Arm	EF	EP	R	Won
Simulator-calibrated	deployment pattern $A^{\prime}$ 0.15	3 / 9	8 / 16	65 / 61	yes
	$A^{\prime}$ 0.25	3 / 10	8 / 20	65 / 66	no
	$A^{\prime}$ soft	3 / 8	8 / 11	65 / 70	no
	law only	3 / 2	8 / 2	65 / 63	no
	filter only	3 / 5	8 / 7	65 / 65	no
	stock planner	3 / 90	8 / 223	65 / 35	yes

TABLE IV: Pooled counts of the proposed stack (M) against every other arm, M / arm, and the two-requirement criterion. EF entry flights, EP entry episodes, R goals reached.

Regime	Stratum	$n$	EF fewer / equal / more	EP fewer / equal / more	R
Simulator-calibrated	clear 0.500	24	1 / 22 / 1	1 / 22 / 1	13 / 21
	clear 0.400	24	1 / 21 / 2	1 / 21 / 2	19 / 13
	fog 0.500	21	3 / 18 / 0	3 / 18 / 0	14 / 14
	fog 0.400	24	4 / 20 / 0	4 / 20 / 0	19 / 13
	pooled	93	9 / 81 / 3	9 / 81 / 3	65 / 61
Transferred	clear 0.500	24	2 / 21 / 1	3 / 20 / 1	15 / 20

TABLE V: Keepout entries cell by cell, the proposed stack against the deployment pattern on the matched cells of each regime. Fewer, equal and more count the cells in which the stack has fewer, as many or more entry flights (EF) or episodes (EP) than the deployment pattern; R counts goals reached, stack / deployment pattern.

Regime	Arm	Flights	Start stalls	Livelocks	Flips	Re-flies	Crashes
Simulator-calibrated	full stack	93	14	6	0	0	0
	deployment pattern $A^{\prime}$ 0.15	93	0	0	1	0	0
	$A^{\prime}$ 0.25	93	0	0	2	0	0
	$A^{\prime}$ soft	93	0	0	1	0	1
	law only	93	0	0	0	0	0
	filter only	93	0	0	0	0	1

TABLE VI: Flight health per arm, pooled on the regime’s cells, flights, start stalls, route livelocks, flips, re-flies and in-mission planner crashes, and the claim rows of the proposed stack against the deployment pattern by cell set of the simulator-calibrated regime (host of the flown cell) and on the same cell sets of the transferred regime. EF entry flights, R goals reached, M /

A^{\prime}

Fig. 1: Three illustrative evaluation cells at fog 0.400 in the simulator-calibrated regime, T17, T20 and T22 being start poses and r5 and r6 replicates, selected by one stated outcome criterion, the deployment pattern A ′ 0.15 entered the keepout in force and the proposed stack M reached the goal and made no entry, and not as representative samples of all flights. Grey, true walls. Light tone, the static keepout. Saturated tone, advance cells in force at some epoch of either flight, which need not be in force at one instant. Black, M. Blue, A ′ 0.15. Dots ringed per metre, entry samples. Triangle, start. Star, goal. In T22 r5 the two A ′ entries lie above the plotted height band and no coloured cell represents them, and the slow, backtracking approach of M to the goal there, 79 s and 41 m of path for the last 15 m, is an execution limitation of the planning layer and not a plotting artefact.

Safe Robot Navigation

Safe robot navigation develops methods for reaching goals while avoiding collisions and managing uncertainty, dynamic obstacles, and environmental hazards.

See all 140 papers

Aug 15, 2026

SCORE: Shape-Conforming Regions for Flight in Enclosed, Degraded Environments

Autonomous UAVs enter enclosed environments such as caves and collapsed structures that confine the vehicle and degrade perception. Conformal prediction provides a distribution-free guarantee by calibrating how far an obstacle keep-out must expand to absorb perception error at a target coverage level. However, existing keep-out regions use convex primitives whose bulges consume narrow passages and grow as perception degrades. Our main contribution defines the nonconformity score on a signed distance field (SDF). This produces a non-convex keep-out that tightly follows obstacle geometry and avoids the unnecessary bulging of equal-margin convex regions. Two supporting components keep this geometry usable as perception degrades. First, a voxelwise union of complementary sensor observations certifies voxels that any single sensor misses. Second, the margin around the obstacle adapts to measured visibility without weather labels or the online ground-truth feedback that single-pass flight cannot provide. Results on real subterranean data show that the resulting distribution-free, shape-conforming keep-out retains more usable free space than convex baselines at the same certified coverage, and produces safer closed-loop flight.

Eric Minwoo Kim, Jong-Kook Kim

School of Electrical Engineering, Korea University, Seoul, Republic of Korea

UAV Navigation Robotics Robot Motion Planning Conformal Prediction

Ask PDF

Aug 27, 2026

Barrier Function Conformal Safety Clearance Certification with CVaR for Driving Trajectory Selection

Autonomous driving motion planners generate and select candidate trajectories while accounting for interactions with surrounding agents. However, these evaluations do not certify the actual safety clearance of the selected trajectory. The framework evaluates the trajectory selected by ant planners and calibrates the gap between its plan time margin and realized safety clearance. A differentiable separating axis barrier margin deterministically lower bounds exact signed oriented-bounding-box (OBB) safety clearance, connecting the statistical certificate to safety margin. At plan time, the margin is evaluated using either a nominal prediction and sampled lower tail Conditional Value-at-Risk (CVaR), while post-selection conformal calibration over exchangeable drive sessions absorbs prediction and sampling errors. Conformal calibration provides statistical validity independently of predictor correctness. The method is evaluated on a frozen 300 session nuPlan study using native Predictive Driver Model (PDM) Closed loop proposals. At 10% target miscoverage, sampled lower CVaR reduces the conformal correction from 1.43m to 0.03m and increases the rate of nonnegative safety clearance certificates from 68.7% to 87.3%. Across all evaluated statistics, exact-clearance coverage remains above the 90% target at 93.3--96.7%.

Pei Yu Chang, Qadeer Ahmed

Department of Mechanical and Aerospace Engineering,The Ohio State University, Columbus, OH 43212 USA

Autonomous Driving Safety Evaluation Autonomous Driving Planning Conformal Prediction

Ask PDF

May 5, 2026

Differentiable Optimization Layered Safety-Critical Control for Risk-Aware Navigation via Conformal Prediction

Risk-aware navigation in unknown environments is a fundamental challenge for autonomous vehicles operating in complex urban systems. To address this issue, this paper presents a differentiable optimization layered safety-critical control method based on conformal prediction. First, to handle uncertainties arising from sensor noise, the conformal prediction method is employed to generate risk-aware obstacle ellipsoids around an elliptical-shaped robot. Second, two nested differentiable optimization layers are introduced to build the control barrier functions for obstacle avoidance and feasibility guarantee, respectively. Then, a quadratic program based safety-critical control law is proposed to integrate the above control barrier function constraints as well as input constraints. In the end, the effectiveness of the proposed framework is demonstrated through numerical simulations.

Jinyang Dong, Shizhen Wu, Yongchun Fang

Institute of Robotics and Automatic Information Systems, College of Artificial Intelligence, Nankai University, Tianjin 300071, China · Academy for Advanced Interdisciplinary Studies, Nankai University, Tianjin 300071, China · Institute of Intelligence Technology and Robotic Systems, Shenzhen Research Institute of Nankai University, Shenzhen 518083, China

Safety-Critical Control Control Barrier Functions Safe Robot Navigation Robot Safety Differentiable Optimization

Ask PDF

Abstract

Safe Robot Navigation

Explore similar work

SCORE: Shape-Conforming Regions for Flight in Enclosed, Degraded Environments

Barrier Function Conformal Safety Clearance Certification with CVaR for Driving Trajectory Selection

Differentiable Optimization Layered Safety-Critical Control for Risk-Aware Navigation via Conformal Prediction