PaperlayerView exact revision on arXiv

Source-linked AI summary

Learning to Optimize UAV Path Planning for Data Sensing in Wireless Sensor Networks

Sijie Ma, Zeyuan Ma, Weijia Cao, Yue-Jiao Gong, Lingling Ma, Zhiyang Huang, Jun Zhang

arXiv:2609.16629v1cs.ROcs.NE

TL;DR

UAV path planning for WSN data collection must contend with handcrafted optimizers and simplified environmental models. LAMDE combines meta-learned optimizer configuration with landscape-aware training augmentation and flexible encoding; experiments report state-of-the-art performance on the tested complex UAV tasks.

  • Problem

    Existing evolutionary methods simplify obstacles, fix hover-point counts, and depend on handcrafted optimizers, while realistic task constraints complicate planning.

    Source: §1 · pp. 1–2

  • Method

    LAMDE meta-learns low-level optimizer settings and combines landscape-aware training-data augmentation with variable-length encoding.

    Source: §1 · p. 2

  • Results

    LAMDE achieves state-of-the-art performance on the proposed UAV tasks compared with traditional and learning-based evolutionary optimizers.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The paper presents LAMDE as both a solver for challenging UAV planning tasks and a practical guideline for developing MetaBBO for realistic problems.

    Source: §6 · p. 13

    The conclusion identifies planning in noisy environments and multi-UAV co-evolution as future directions.

    Source: §6 · p. 13

Abstract

from arXiv · show

UAVs have emerged as highly flexible platforms for data sensing in Wireless Sensor Networks (WSNs). Path planning for UAVs in such tasks plays a key role to assure remote sensing effectiveness and friendly energy consumption. However, existing approaches show two key limitations: i) they are primarily hand-crafted with certain design biases that harm adaptation on unseen tasks. ii) they predominantly assume idealized spatial complexities of actual environments through simplified simulation, causing them to underperform during real-world deployment. In this paper, we propose a novel learning-assisted planning framework, termed Landscape-Aware Meta Differential Evolution (LAMDE), to tackle the mentioned limitations. The major contributions come from the following aspects. We first re-formulate such UAV path planning problem to embrace challenging constraints. To efficiently navigate this highly constrained space, we propose a bi-level learning to optimize approach, where the meta-level is a trainable algorithm configuration policy that meta-learns an adaptable planning strategy for low-level planning algorithm. To address the potential training data scarcity and distribution shift in real-world environments, we introduce a landscape-aware automatic augmentation scheme that enriches training data. At the low-level, a Differential Evolution algorithm is deployed for solving the path planning tasks. To enhance the solving flexibility, we further design a variable-length encoding strategy that dynamically prunes redundant hover points and optimizes continuous flight parameters concurrently within a unified search space. Based on all proposed designs, we meta-train LAMDE and compare it with representative baselines. Comprehensive experiments demonstrate that LAMDE achieves state-of-the-art performance on the tested complex UAV path planning tasks in WSN data collection scenarios.

1 Introduction

LAMDE targets limitations in realistic UAV path planning by combining a trainable optimizer strategy with problem reformulation and flexible search encoding. Experiments report state-of-the-art performance on the proposed UAV tasks against traditional and learning-based evolutionary optimizers.

1 Introduction

  • Existing evolutionary approaches simplify environments, fix hover-point counts, and rely on hand-crafted optimizers, limiting realistic adaptation and generalization.These approaches may require labor-intensive manual mechanisms and sensitive hyperparameters.Source: §1 · pp. 1–2
  • LAMDE uses a meta-level reinforcement-learning configuration policy to dynamically configure a low-level evolutionary optimizer for end-to-end planning.The policy is trained over a problem distribution to improve expected planning performance.Source: §1 · p. 2
  • The reformulated trajectory model explicitly includes heterogeneous building geometries and customized no-fly zones.Source: §1 · p. 2
  • Variable-length encoding flexibly represents solutions and automatically prunes redundant trajectory points during low-level optimization.The design is intended to improve optimization flexibility and convergence.Source: §1 · p. 2
  • Landscape-aware augmentation matches target-task landscapes to synthetic problems, addressing scarce real-world training data and training–test distribution alignment.Source: §1 · p. 2
  • LAMDE achieves state-of-the-art performance on the proposed UAV tasks compared with traditional and learning-based evolutionary optimizers.Ablations and in-depth analysis also underscore the effectiveness of the proposed designs.Source: §1 · p. 2

2 Related Work

Prior UAV path-planning work spans discrete routing and continuous trajectory optimization, but often represents environments coarsely and uses rigid, hand-designed search. MetaBBO offers adaptive optimizer control, while distribution shift between synthetic training tasks and realistic tests remains a practical concern.

2.1 UAV Path Planning

  • UAV planning evolved from discrete TSP-like sensor visiting to continuous spatial paths, yet many models omit complex building geometries and no-fly zones.Obstacle-free assumptions and coarse grids can mismatch real deployment constraints.Source: §2.1 · p. 2
  • EC methods include GAs, ACO, PSO, and DE, but fixed hover-point counts and handcrafted tuning restrict adaptation across complex missions.The cited methods include specialized routing heuristics, energy-focused routes, and continuous optimizers.Source: §2.1 · pp. 2–3

2.2 Meta-Black-Box Optimization

  • MetaBBO trains a meta-level agent to guide low-level optimizers, with paradigms including algorithm selection, configuration, solution manipulation, and generation.Training approaches include reinforcement learning, supervised learning, neuroevolution, and LLM-based in-context learning [35].Source: §2.2 · p. 3
  • MetaBBO methods often assume synthetic training tasks align with realistic tests, although real-world task distributions may be difficult to characterize or divergent.LAMDE is motivated by this distribution-shift challenge and introduces landscape-aware data preparation.Source: §2.2 · p. 3

3 Problem Definition and Modeling

The model plans a 2D UAV route with hovering-based data collection, adds realistic obstacle geometry and collision-risk measures, and minimizes flight and hover time subject to operational, communication, data, and safety constraints.

3.1 UAV Path Planning in WSN Data Collection

  • The UAV follows a 2D route from a prescribed start to destination, collecting sensor data while hovering and flying between path points.Flight segments use a trapezoidal velocity profile; collected sensor data is aggregated from hover duration and transmission rate.Source: §3.1 · pp. 3–4

3.2 Environmental Modeling with Obstacles

  • The environment model includes circular and arbitrarily oriented rectangular obstacles, plus polygonal no-fly zones, to represent physical and regulatory constraints.The obstacle sets are described as heterogeneous static constraints that introduce non-convexity.Source: §3.2 · p. 4

3.3 Modeling of Collision Risk

  • Collision risk Lvio(P) is the sum of trajectory-overlap lengths across rectangular obstacles, circular obstacles, and no-fly zones.Source: §3.3 · p. 4

3.4 Optimization Model

  • The optimization minimizes weighted flight and hovering time while enforcing path bounds, duration limits, communication bandwidth, per-sensor data requirements, and zero collision penalty.The decision variables are the path P and hover durations Thov; the final constraint nullifies Lvio(P).Source: §3.4 · p. 5

4 Methodology

LAMDE couples a low-level Differential Evolution optimizer with a meta-level policy that configures its search, and trains that policy on landscape-matched synthetic problems. Its methods address constrained path planning through flexible hover-point encoding, constraint-aware optimization, and automated training-data construction.

4.1 Low-Level Optimizer

  • The low-level Differential Evolution optimizer handles constraint violations with penalties and uses their aggregate both for selection pressure and as a meta-policy state feature.The penalty for obstacle overlap is normalized to rebalance its scale relative to communication-bandwidth and data-collection violations.Source: §4.1 · pp. 5–6
  • Variable-length encoding adds an activation value to each candidate hover point, allowing threshold-based pruning of inactive points from the decoded flight plan.The threshold is adjustable; a larger threshold yields a sparser plan.Source: §4.1 · p. 5

4.2 Meta-Level Policy

  • At each generation, the meta-level policy conditions on the optimization state and configures DE mutation and crossover operators and their hyperparameters.The state includes the population, objective values, constraint violations, and search progress; mantissa-exponent decomposition stabilizes evaluation magnitudes across tasks.Source: §4.1 · p. 6, §4.2 · p. 6
  • The reward encourages reducing constraint violations early, gives a bonus for entering the feasible region, and then focuses on objective improvement.The policy uses a predefined feasibility tolerance to distinguish entry into the feasible region.Source: §4.2 · p. 7

4.3 Training

  • Landscape-aware training matches target-task objective and constraint features to synthetic instances, pairing matched problems to construct the augmented training set.For each pair, the objective evaluation comes from one instance and the constraint-violation measure from its matched instance.Source: §4.3 · p. 8
  • LAMDE is meta-trained with PPO on the constructed problem set, where DE evolves a population and the policy receives state, action, and reward feedback at each generation.The trained policy is used to solve unseen tasks in a zero-shot manner.Source: §4.3 · p. 8

5 Experimental Results

LAMDE is evaluated on constrained UAV path-planning tasks through benchmark comparisons, ablations, and feasibility analysis. It generally outperforms the tested baselines, while variable-length encoding and landscape-aware training-data preparation show benefits in the reported studies.

5.1 Experimental Setup

  • The Dtest benchmark contains 16 WSN data-collection settings spanning two map sizes, increasing obstacle densities, and a 10,000-evaluation limit per instance.Eight representative evolutionary and MetaBBO optimizers are compared, with penalty handling applied to several methods not designed for constrained optimization.Source: §5.1 · pp. 9–10

5.2 Systematic Performance Comparison

  • LAMDE significantly outperforms the compared baselines on almost all 16 instances and leads the MetaBBO baselines, though CMA-ES slightly outperforms it on T4 and T11.The comparison aggregates results over 25 independent runs using a penalty objective; the authors also report LAMDE ranks first consistently among the tested methods.Source: §5.2 · p. 10
  • Across the four visualized maps, LAMDE’s best trajectories are more concise and smoother, while the three baselines frequently produce irregular paths with redundant detours.The authors attribute this pattern to the generalized meta-level policy and low-level variable-length encoding that can prune unnecessary hover points.Source: §5.2 · pp. 10–11, Figure 2 · p. 10

5.3 Ablation Studies

  • LAMDE achieves higher and more stable returns than fixed-encoding LAMDE-F across all 16 instances, while LAMDE-F still slightly outperforms the selected baselines.LAMDE-F enforces Imax hover points, whereas the variable-length strategy adaptively prunes redundant trajectory points.Source: §5.3 · p. 11, Figure 3 · p. 11
  • LAMDE-R performs worst overall, while LAMDE-E, LAMDE-L, and LAMDE-O each remain inferior to LAMDE, which achieves the best performance across the 16 instances.The authors link the result to global centroid matching of synthetic tasks and conclude that both learned features and holistic matching matter for zero-shot generalization.Source: §5.3 · pp. 11–12, Table 2 · p. 11

5.4 In-depth Analysis

  • At Imax = 100, LAMDE maintains a feasible rate above 0.6, whereas LAMDE-F fails to produce feasible solutions once Imax ≥70.The analysis counts feasible solutions in the last-generation population across 16 tasks and 25 runs; LAMDE-F feasibility deteriorates as the maximum dimension grows from 20 to 100.Source: §5.4 · p. 12, Figure 4 · p. 12

6 Conclusion

The conclusion presents LAMDE as a MetaBBO framework for realistic UAV path planning in WSN data collection, combining landscape-aware training-data preparation with variable-length encoding. The reported experiments show consistent gains over classical and learning-based baselines, while 3D settings, noisy online planning, and multi-UAV co-evolution remain potential future directions.

6 Conclusion

  • LAMDE addresses UAV path planning for WSN data collection with challenging obstacle layouts and running constraints.The framework reformulates the optimization model around city-level obstacles and uses landscape-aware training-data preparation and variable-length encoding.Source: §6 · p. 12
  • Experiments on diverse UAV missions show that LAMDE consistently outperforms the tested classical and learning-based baselines.The conclusion also reports that in-depth analysis supports the effectiveness of the proposed designs.Source: §6 · p. 12
  • The authors identify 3D environments, online planning in noisy environments, and multi-UAV co-evolution as potential future work.Source: §6 · p. 13

Related papers

Loading 2609.16629v1…