Residual-state accounting targets leftover inspection opportunity in UAV route planning

A receding-horizon planner described as RAWA-RHP tackles a problem that arises after a model-feasible route prefix has already been constructed for an unmanned aerial inspection mission: deciding how much of the remaining inspection opportunity can still be exploited. Its central idea is a residual-state formulation that separates travel and return commitments from adaptive-reserve feasibility slack and risk-proxy headroom, allowing a specialised insertion procedure, RR-CMP, to assess additional targets within the planning model’s own feasibility checks rather than outside them.

The planner operates at the high-level waypoint-selection layer and returns the first action of a model-feasible prefix, replanning after execution. Wind-aware energy budgeting relies on an empirical q90 edge statistic computed from 32 gust samples, an estimate whose q90 quantile is governed by only a few upper order statistics. An adaptive reserve and explicit return accounting sit alongside this tail-budget baseline, so candidate insertions load the remaining budget quantities differently depending on where they are placed in the prefix.

Controlled comparisons support the insertion step more strongly than the full pricing rule. Relative to no packing, Full RR-CMP increased mean safe weighted coverage, or SWC, by 0.1640, a metric in which any included failure sets an episode to zero. Its difference from simple feasible insertion, SFI, was 0.0045 SWC, with an interval spanning zero and a Holm-adjusted p-value of 0.621. The study therefore does not establish a statistically reliable additional advantage for the elaborate gain expression over plain feasible insertion, nor does it isolate the causal effect of the residual-state representation alone.

RAWA-RHP also obtained higher SWC than Adaptive-FairALNS under the evaluated budget controls, Native, EqTime and EqEval. The controls answer different questions: EqTime provides an equal wall-clock allowance rather than equal active computation, and EqEval retained only 0.879 of the RAWA-RHP mean primitive count, a 12.1 percent comparator shortfall that could favour RAWA-RHP. Because primitive costs also differ, EqEval is an evaluation-count control with known residual bias rather than a strict fairness proof.

A deterministic mixed-integer programme covered easy 8-to-16-target fixed-cost instances and certified all tested cases. At 16 targets, the mean objective gap between RAWA-RHP and SFI was about 0.0007, with mean gaps of 0.0094 for RAWA-RHP and 0.0087 for SFI; the objective values 32.10 and 32.12 differ by 0.02. The study notes that these small certified instances cannot characterise large-scale approximation loss or the global optimality of the stochastic planner.

Evidence beyond the controlled suite is more mixed. An exploratory generalisation matrix of 30 paired scenarios favoured SFI, with RAWA-RHP at 0.7963 SWC against 0.8259 for SFI, and most subgroup intervals were inconclusive. In 3D execution, an uncalibrated transfer stress test retained evaluated-failure rates of 83.3–85.0 percent across planners. Independent nominal-only calibration improved nominal return behaviour, yet strong and correlated gusts continued to produce substantial battery and return failures, which the authors attribute to limitations in the current energy-model and controller calibration rather than evidence of operational robustness.

The intended deployment is a high-level waypoint-selection service above a flight controller, state estimator and local collision-avoidance layer. A real integration would supply a time-stamped wind forecast, a battery state and health estimate, a vehicle-specific edge-energy model, localisation covariance and a clearance map, while the flight controller tracks only the next accepted waypoint and could veto it. At runtime, uncertainty outside the calibration envelope should trigger a conservative mode that suspends further inspection, retains or increases reserve, commands return or hold through the flight manager, and requests replanning after state validation. If the 1.75-second internal search cap expires, the implementation returns its best already checked prefix, and if no such prefix exists it returns rather than forcing a target.

Before transfer, the authors specify vehicle-specific calibration priorities. Energy coefficients should be fitted to logged voltage, current, ground speed, airspeed, payload and wind over representative manoeuvres, with held-out evaluation of the q90 residual. The reserve floor should combine regulatory or operational return policy, battery ageing, landing energy and an upper confidence bound on model error rather than reusing the 80 simulator units. The clearance buffer should at least sum vehicle radius, required physical standoff, the tracking-error bound, localisation and map uncertainty, and controller stopping distance. Transmission corridors, bridge decks and wind turbines should not share one unmodified parameter set, since each geometry demands asset-specific modelling of conductor standoff, multipath and occlusion, or blade motion and local aerodynamic effects.

The limitations remain substantial. The evidence is simulation-based; the 3D model adds dynamics, tracking, correlated disturbance, execution-energy mismatch and executed clearance, but it is neither hardware-in-the-loop nor real flight. The dimensionless coefficients are operating-range choices rather than identification of a named vehicle, and the clearance-risk proxy assumes an independent product over segments that the correlated-execution assessment does not support as a probability; explicit planner proxy caps are disabled in the reported configuration, so clearance behaviour is governed by proxy-aware ranking, reserve effects and evaluated execution checks rather than a chance constraint. Transfer of latency from fixed-core CPU emulation to an onboard processor is uncertain because cache, thermal throttling, middleware and concurrent perception loads differ, and no real-flight, recorded-wind or hardware-platform evidence was available.

The resulting design lesson is that remaining inspection opportunity should be made explicit and the contribution of each improvement stage tested separately, rather than assuming that additional scoring complexity will yield broader gains. The controlled results support examining insertion and pricing complexity independently: the residual-state accounting demonstrably helps relative to no packing, while the full pricing rule’s advantage over simple feasible insertion was not statistically reliable in this planning setting.

Subject of Research: RAWA-RHP: Residual-state constrained insertion for wind-affected UAV infrastructure inspection

Article Title: RAWA-RHP: Residual-state constrained insertion for wind-affected UAV infrastructure inspection

Article References: Li, H., Wei, J., Yao, D., Zhao, Y., Li, Y., Mai, Z., Yu, J., Lu, B., Im, S.-K., Yang, X., & Wang, Y. (2026). RAWA-RHP: Residual-state constrained insertion for wind-affected UAV infrastructure inspection. Results in Engineering, 32, Article 113481. https://doi.org/10.1016/j.rineng.2026.113481

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.113481

Keywords: RAWA-RHP, Residual-state, constrained, insertion, wind-affected, infrastructure, inspection, scientific research