Where Should Risk Live? A Boundary Analysis of Risk-Aware Coverage Exploration for an Ackermann Robot

Risk-aware frontier selection trades exploration efficiency for driving safety, yet it is rarely tested whether the risk term changes any decision under a real platform's constraints — or which layer should carry the safety requirement instead. I answer both on a 0.52 m Ackermann vehicle in Gazebo / ROS 1 with a force-controlled execution layer, whose dynamics I identify from open-loop steps (τ = 0.216 s), from which all parameters follow. With truth-referenced metrics (397.0 m² free space, three repetitions), information-gain, boustrophedon and random policies cover 93.6%, 79.0% and 72.1%. Sweeping the selection-layer risk weight β across four scenes whose candidate risk spread Δρ grows from 0.05 to 0.80 produces no measurable effect. I derive the condition β·Δρ > gain advantage, and explain the degeneracy geometrically: with a 0.65 m turning radius, risky corridors are too narrow to traverse. Moving the constraint into the execution layer (a big-M penalty in a sampling MPC) raises minimum clearance from 0.18 to 0.25 m on point-to-point traversals in three geometries (36 runs) at ≤0.5% cost, with a reproducible +0.05 m bias, but fails when reused in the exploration loop. Feeding the measured passable width into the policy-layer graph search closes the loop: failures fall by 96% and cumulative risk by 93% at unchanged coverage. The results consolidate into a margin rule — clearance minus threshold must exceed tracking error — placing the executable limit of a 0.25 m constraint at 1.26 m. All results are from simulation; hardware transfer and ROS 2 migration are future work.

Changes in v2: Yule Wang added as co-author (supervision and manuscript review), with the affiliation given at school level. The abstract was shortened to 250 words to meet journal limits and the statements section was moved after the references. Data, analyses and conclusions are unchanged.