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

Risk-aware frontier selection is widely used to balance exploration efficiency against driving safety in autonomous mobile robots. Yet two questions are rarely asked: does such a risk term actually change the robot's decisions under a real platform's constraints, and, if it does not, which layer of the system should carry the safety requirement instead? I answer both on a 0.52 m Ackermann vehicle built in Gazebo / ROS 1 with a force-controlled execution layer, so that servo compliance and latency are present, and whose chassis dynamics are identified from open-loop step tests (time constant 0.216 s; steering delay 0.17 s; 2.3 degree asymmetry). Using truth-referenced metrics (coverage against ground-truth reachable free space, 397.0 square metres; three repetitions per condition), I compare three coverage policies, which obtain 93.6, 79.0 and 72.1 percent coverage. Sweeping the selection-layer risk weight across four scenes whose candidate risk spread grows from 0.05 to 0.80 produces no measurable effect; I derive the criterion that the weight times the risk spread must exceed the gain advantage, and explain the degeneracy geometrically via the 0.65 m minimum turning radius. Moving the constraint into the execution layer, as a big-M clearance penalty inside a sampling MPC, raises minimum clearance from 0.18 m to 0.25 m on point-to-point traversals (36 runs, at most 0.5 percent path length cost) and shows a reproducible +0.05 m conservative bias. Reused in the full exploration loop the constraint fails; the cause is the planning layer handing locally infeasible paths to the controller. Feeding the measured passable width into the policy-layer graph search closes the loop, cutting failures by 96 percent and cumulative risk by 93 percent at unchanged coverage. The results consolidate into a margin rule: available clearance minus the threshold must exceed the tracking error, which places the executable limit at about 1.26 m of clear width on this platform.