METHOD FIGURE
The proposed method separates task-focused wind learning from monitoring.
(a) Given the obstacle map, current robot state and clarity map, maintain a
spatial-only POAM wind belief. The actual wind is unknown to the method.
(b) The existing online geometric initializer generates several collision-free
paths to high-value sensing destinations. (c) For each path, its endpoint value
V_k is the sensor-weighted clarity deficit predicted without sensing at H/4,
H/2 and H. That value weights a compact tube around the entire transit path.
The resulting nonnegative weights m_j are normalized to sum to one. The tube
radius is r = 0.25 m; psi(s) = (1-s)^4(1+4s) on 0 <= s <= 1 and zero outside.
This is a task-relevance proxy, not an exact value-of-information calculation.
(d) For candidate measurement z, the per-component expected reduction is
Delta V_m(z) = sum_j m_j C(p_j,z)^2 / (sigma^2(z) + sigma_n^2), using the full
posterior covariance, including its conditional residual. The common factor
of two for the two wind components does not change the ranking. Candidate
waypoints are ranked by Delta V_m(z)/(1 + L(z)/ell_0), with ell_0 = 1 m and
geometric travel length L. Persistent goals, feasible-prefix screening and
stalled-route recovery remain part of the existing navigation policy. RPCBF
filters the executed control. Only actual local noisy measurements update POAM.
(e) Routes and the task mask refresh every 2 s. Learning ends when the
task-weighted posterior vector RMS U_m <= 1.10 m/s. RMSE against the simulated
truth is a development diagnostic, never the handoff trigger. The illustrative
center-start Scenario 4 recording switches at 124 s. Panel (f) uses its first
recorded post-handoff forecast spanning both monitoring regions, at 164 s.
(f) The same geometric alternatives, a hold and a shifted previous reference
initialize independent MBD chains. Each candidate is evaluated under spatially
coherent samples from the current POAM posterior, with the double-integrator
robot, relative-air-velocity force, tracking controller, relaxed planning RPCBF
and full clarity dynamics. The cost is expected accumulated clarity deficit;
there is no added information reward in MBD. The belief is frozen within a
planning call; hypothetical observations are never assimilated. The controller
executes 5 s before replanning a 15 s horizon, collecting actual measurements
and continuing POAM updates. Execution uses the existing full RPCBF with
infeasibility logging and fallback; this is safety-aware, not a universal safety
guarantee. The two-stage method shown does not add a new return-to-survey mode.

PERFORMANCE FIGURE
Mean positive clarity shortfall over all free cells, D(t) = mean_j
max(q_target,j - q_j(t), 0), plotted from the retained full-resolution samples.
Panel order: Scenario 1, Scenario 2, Scenario 4, Scenario 3. The curves end at
their shared comparison windows: 532, 560, 544 and 514 s, respectively, or at
recorded contact. The former vertical comparison line is now the right edge
of each plot. No data after that line are displayed; no continuation is added
after a contact. Diamonds mark the end of a dedicated learning stage; crosses
mark contact. All plots share the same symmetric-log y-axis: linear from zero
to 0.02 and logarithmic above 0.02. No curve smoothing or temporal averaging is
applied. Scenario 4 uses its declared center initial condition. All four scenarios
include all five recorded methods. Targets and
initial clarity differ between scenarios, so compare methods within each
panel; absolute deficit levels are not a normalized cross-scenario ranking.
These are individual recorded runs, not seed averages or confidence intervals.

Baseline 1: still-air MBD, no RPCBF.
Baseline 2: still-air MBD with learned-belief tracking and RPCBF at execution.
Baseline 3: disturbance-aware MBD, online learning, no initial survey.
Baseline 4: global uncertainty reduction, followed by disturbance-aware MBD.
Proposed: clarity-valued candidate-path learning, followed by the same MBD.
