PaperlayerView exact revision on arXiv

Source-linked AI summary

Large-Scale Tunnel Air-Ground Collaboration With FLISP: Fast LiDAR-IMU Synchronized Path Planner

Fenghe Guo, Runjie Shen, Chenyang Sun, Junrui Zhang, Quanxi Zhan, Yongchun Wang, Junjie Zhang

arXiv:2606.25393v2cs.RO

TL;DR

Hydropower tunnel inspection requires safer, more effective coordination than existing single-robot and map-dependent approaches provide. FLISP uses a mapless, UGV-centric LiDAR-IMU framework to generate synchronized UGV-UAV paths, achieving a 100% success rate in a 1.2 km operational tunnel.

  • Problem

    Synchronous UGV-UAV tunnel planning remains scarce, while map-dependent systems are vulnerable to SLAM drift in featureless tunnel geometry.

  • Method

    FLISP centrally processes a single UGV-mounted LiDAR-IMU stream to generate synchronized UGV-UAV trajectories through a hierarchical, mapless planning pipeline.

  • Results

    100% success rate was achieved during benchmarks and deployment in a 1.2 km operational tunnel, outperforming map-based baselines in efficiency and real-time control.

  • Takeaways & Limitations

    FLISP provides a practical planning framework for automated inspection in large-scale, geometrically degenerate hydropower infrastructure.

Abstract

from arXiv · show

Hydropower tunnel inspection is critical for infrastructure integrity yet remains inefficient and hazardous using manual methods. We propose FLISP (Fast LiDAR-IMU Synchronized Path Planner), a mapless planning framework for cooperative UGV-UAV inspection. Unlike traditional map-based paradigms, FLISP features three core contributions: (1) a unified architecture where a single UGV-mounted LiDAR-IMU suite drives synchronized path generation for both platforms; (2) platform-specific solvers utilizing an enhanced Firefly Algorithm for UGV obstacle avoidance and a dynamic iterative optimizer for UAV flight; and (3) a hierarchical refinement strategy ensuring kinematic feasibility without state estimation drift. Benchmarks in a 1.2 km operational tunnel demonstrate that FLISP circumvents structural bottlenecks of map-based methods, eliminating map rasterization overhead (Fast-LIO2 + A*) and sampling instability (LIO-SAM + RRT*). FLISP achieves a 100% success rate with 7 ms latency, representing a 7-fold speedup over grid-based and a three-order-of-magnitude improvement over sampling-based baselines. Validated in operational hydropower tunnels, this approach offers a scalable solution for robotic inspection in feature-degraded linear infrastructure. A demonstration video is available at https://youtu.be/Y_ezs1PfLJ4, and the code at https://github.com/ArchibaldGuo/FLISP.git.

I. INTRODUCTION … D. METHODS IN TUNNEL ENVIRONMENTS

FLISP addresses the hazards and limitations of tunnel inspection by enabling mapless, synchronized UGV-UAV planning from a single UGV-mounted LiDAR-IMU suite. It distinguishes platform-specific path planning and kinematic refinement from prior map-dependent, single-platform, or homogeneous multi-robot approaches.

  • I. INTRODUCTION: FLISP targets tunnel-inspection challenges that defeat single robots, including inaccessible tunnel crowns, limited UAV endurance, GPS denial, repetitive textures, uneven surfaces, and multi-agent collision avoidance.Manual inspection is labor-intensive, hazardous, and vulnerable to subjective errors –.
  • I. INTRODUCTION: FLISP enables synchronized, collision-free UGV-UAV path planning from local geometry captured by one UGV-mounted LiDAR-IMU suite in degenerate tunnel environments.Its UGV-centric, vertically integrated architecture reduces computational load and avoids SLAM drift accumulation.
  • I. INTRODUCTION: FLISP combines platform-specific obstacle-avoidance strategies with multi-level optimization for UGVs on uneven curved floors and UAVs near fixed infrastructure such as sluice gates.The framework is designed to maintain real-time planning while satisfying distinct platform constraints.
  • A. CLASSIC PATH PLANNING ALGORITHMS: Classical planning separates global methods based on prior maps from local methods driven by real-time perception, including A*, Dijkstra, RRT, APF, DWA, and deep learning approaches –.Global planners optimize reference trajectories against topological, energy, and safety constraints, whereas local planners generate online trajectories with sensor fusion and kinematic feasibility.
  • B. PATH PLANNING FOR MULTI-ROBOT SYSTEMS: Learning-based multi-robot planners, including SAC and MAPPO with experience replay, and heuristic methods such as hybrid-state A* with conflict-based search, generally target simulated, structured, or homogeneous settings.These assumptions limit generalization to unmapped, non-convex tunnel geometries and heterogeneous kinematic constraints.
  • C. PATH PLANNING FOR HETEROGENEOUS MULTI-ROBOT SYSTEMS: Heterogeneous UAV-UGV systems commonly combine UAV global mapping with UGV local sensing, while centralized nodes or mediator agents coordinate data and resolve conflicts,,.LiDAR-based collaborative planning with tether optimization incurs computational overhead that can produce stepwise motion and restrict real-time agility.
  • D. METHODS IN TUNNEL ENVIRONMENTS: Tunnel-planning methods remain challenged by uneven terrain, GNSS denial, geometric uniformity, and unreliable perception in low-light or curved sections, while prior approaches address only selected UAV or UGV estimation problems,.These limitations motivate a unified solution for degraded linear infrastructure.
  • D. METHODS IN TUNNEL ENVIRONMENTS: Existing synchronous UGV-UAV frameworks depend on drift-prone global SLAM, while subterranean systems – and single-platform mine planners, often assume flat traversability or isolate platforms; FLISP instead performs simultaneous mapless planning with platform-specific constraints.A SLAM failure in featureless geometry can collapse the collaborative system, and flat-ground assumptions omit heavy-UGV kinematics on curved tunnel floors.

III. SYSTEM ARCHITECTURE … 1) Foundational Corridor Generation

FLISP uses a UGV-centered LiDAR-IMU architecture to generate and control synchronized UGV-UAV trajectories without a globally consistent map. Its foundational UGV corridor planner estimates tunnel orientation, adaptively samples geometry, infers occluded boundaries, and produces a long-horizon centerline for smooth execution.

  • B. UAV AND UGV HARDWARE: The hardware combines a DJI Mavic 3T with a custom GPU-equipped UGV carrying a 64-line navigation LiDAR-IMU, UAV tracking sensors, and wireless infrastructure.The UAV tracking module uses a short-range LiDAR and monocular RGB camera on a two-degree-of-freedom gimbal.
  • C. COMMUNICATION TOPOLOGY: A dual-router wireless bridge connects the UGV, UAV, tunnel entrance, and Ground Control Station for real-time communication over extended operational distances.The UGV creates a local Wi-Fi zone for the UAV while linking wirelessly to a static entrance router connected to the Ground Control Station by Ethernet.
  • D. OPERATIONAL PHILOSOPHY: FLISP collects fixed-rate gimbal-camera imagery online and reconstructs the tunnel offline, requiring constant velocity and a jitter-free trajectory for alignment.This operational strategy avoids reliance on real-time SLAM, whose drift in featureless tunnels can compromise online defect localization.
  • IV. FAST LIDAR-IMU SYNCHRONIZED PATH PLANNER: FLISP coordinates UGV perception, hierarchical UAV planning, and feedback control through a UGV-centered architecture using LiDAR-IMU data and local odometry.The system does not require a globally consistent map, but accurate UGV local odometry remains necessary for short-term trajectory tracking.
  • A. Hierarchical UGV Path Planner: The hierarchical UGV planner consists of corridor generation, refinement, and dynamic and safety constraint smoothing, with the corridor providing a long-horizon geometry for feasible motion.This anticipatory geometry supplies a yaw gradient for gradual micro-adjustments that accommodate slow-response wheel arms and reduce collision risk from sudden steering commands.
  • 1) Foundational Corridor Generation: The foundational corridor stage estimates vehicle orientation relative to tunnel walls, then derives a coarse but directionally correct centerline from raw sensor data.Because six-axis IMUs lack absolute yaw in magnetically complex environments, the planner estimates relative yaw from sampled wall normals and fuses it into the attitude quaternion.
  • 1) Foundational Corridor Generation: Adaptive binning uses yaw-dependent step size to increase path resolution in turns while accelerating planning through straight tunnel sections.Within each bin, the planner extracts leftmost and rightmost boundary points before fitting the corridor geometry.
  • 1) Foundational Corridor Generation: Multi-level boundary fitting bridges LiDAR cavities caused by water reflection, preserving path continuity despite locally fragmented point clouds.The resulting initial UGV path is the midpoint between fitted boundaries, though it may remain noisy before later refinement.

2) Path Robustification and Refinement

The refinement stage converts a noise-sensitive coarse path into a smoother, more reliable trajectory using Bayesian outlier correction. It then projects the corrected path onto the tunnel floor by fitting elevation from the 3D point cloud to follow ground topology.

  • Bayesian path correction: Bayesian refinement improves the initial path’s smoothness and reliability by correcting sensor-noise-induced outliers.The initial path is treated as a coarse blueprint before refinement.
  • Bayesian path correction: Each point’s predicted position averages forward and backward linear interpolations from neighboring points.The method computes separate local predictions before combining them.
  • Bayesian path correction: Measurement errors are modeled with Gaussian distributions using smaller σn for normal points and larger σo for outliers.The error is defined as the difference between actual and predicted positions.
  • Ground-topology projection: The corrected 2D path is projected onto the tunnel floor by fitting elevation from the 3D point cloud, producing a 3D path that follows ground topology.This projection step is illustrated in Fig. 9.

3) Dynamic Safety and Obstacle Avoidance

FLISP handles dynamic obstacles with path-centric safety corridors tailored to each platform, then accelerates detection through path-aligned spatial partitioning. When avoidance is required, an enhanced Firefly Algorithm performs constrained 1D lateral optimization with a dynamically increasing tilt penalty near tunnel walls.

  • Dynamic obstacle detection: FLISP constructs a path-centric safety corridor and checks obstacles only within it, using rectangular volumes for the UGV and cylindrical volumes for the UAV.The detection volume is defined per path segment in the global frame and tailored to each platform’s geometry.
  • Dynamic obstacle detection: A path-aligned 1D grid maps LiDAR points with O(1) indexing, restricting precise geometric checks to occupied cells while preserving accuracy.This broad-phase filter reduces the O(N · M) brute-force burden needed for dense LiDAR frames within a 10 ms control cycle.
  • Local avoidance optimization: Upon obstacle detection, enhanced FA searches laterally in 1D within the tunnel cross-section while enforcing tunnel-boundary and maximum-inclination constraints.The reduced search space accelerates convergence while maintaining collision-free, physically feasible paths for the UGV.
  • Local avoidance optimization: The dynamically weighted cost increases tilt penalties non-linearly near walls, creating a soft barrier against rollover in the curved tunnel geometry.The objective also rewards obstacle clearance while penalizing center deviation and excessive tilt.

4) Final Path Smoothing and Pose Correction

The final stage converts the corrected path into a smooth, executable UGV trajectory. FLISP reduces rollover risk through angle shrinkage, adaptive transition points, orientation interpolation, and quaternion-based transformation.

  • Final Path Smoothing and Pose Correction: FLISP smooths the final path and attenuates measured angles with a shrinkage factor to mitigate rollover risk.This pose correction is applied before generating the executable trajectory.
  • Final Path Smoothing and Pose Correction: The method adaptively determines the number of transition points from the yaw angle’s magnitude.The transition-point count is calculated using Eq. (22).
  • Final Path Smoothing and Pose Correction: Vehicle orientation is linearly interpolated across the transition points, then quaternions transform local positions into the executable UGV path ϕ.The quaternion-based final transformation is defined in Eq. (24).

B. UAV PATH PLANNER · 1) Initial Path Planning · 2) Dynamic Sampling Iterative Path Optimization

The UAV planner hierarchically derives a synchronized aerial path from the refined UGV path, enforcing communication, altitude, and obstacle constraints. It initializes a communication-safe route, then dynamically optimizes waypoint candidates using a multi-objective cost function to produce the final path.

  • B. UAV PATH PLANNER: The UAV planner uses the refined UGV path ϕ as its reference while enforcing aerial communication, altitude, and obstacle constraints for synchronized movement.This hierarchy exploits the UAV’s maneuverability while maintaining coordination with the UGV.
  • 1) Initial Path Planning: The initial UAV path Γinitial is formed by projecting desired-altitude points Ph into communication safety constraints built around the minimum-altitude path Plowest.Plowest is generated from the refined UGV path, while Ph is projected from the tunnel centerline.
  • 2) Dynamic Sampling Iterative Path Optimization: FLISP dynamically samples candidate waypoint points ζi in polar coordinates around each initial-path waypoint to optimize smoothness and obstacle avoidance.The iterative optimization process is summarized in Fig. 12.
  • 2) Dynamic Sampling Iterative Path Optimization: The safety cost J1 penalizes obstacle proximity within a cylindrical corridor, while smoothness cost J2 penalizes deviation from a straight line.Obstacle penalties are quadratic and depend on the obstacle’s distance to the path segment.
  • 2) Dynamic Sampling Iterative Path Optimization: The progress reward J3 favors efficient goal-directed movement, whereas height consistency cost J4 penalizes unnecessary altitude variation to reduce energy use.Together, these terms are combined in the multi-objective cost function used to evaluate candidates.
  • 2) Dynamic Sampling Iterative Path Optimization: The final optimized UAV path ϵ consists of the sequence of candidate points selected by minimizing the total cost.This converts the initial safe path into the optimized aerial trajectory.

C. TRAJECTORY PLANNING AND CONTROLLER … 1) Nominal Planning Robustness

FLISP converts sensor-derived, kinematically aware waypoints into feasible trajectories for both vehicles using lightweight generation and feedforward PD control. In simulation, it was evaluated for nominal tracking, dynamic obstacle avoidance, and runtime efficiency across tunnel geometries, with UGV stability robust to sparse point clouds and noise.

  • C. TRAJECTORY PLANNING AND CONTROLLER: FLISP directly generates high-quality, kinematically aware paths from sensor data, avoiding computationally expensive trajectory-optimization layers and enabling lightweight tracking.This design supports standard trajectory generation and simple controllers.
  • C. TRAJECTORY PLANNING AND CONTROLLER: Discrete FLISP waypoints are converted into continuous, time-parameterized trajectories for both platforms.The UGV uses minimum-jerk quintic-polynomial interpolation, with dynamic time reallocation enforcing velocity and acceleration limits.
  • C. TRAJECTORY PLANNING AND CONTROLLER: The UAV uses non-uniform B-splines with curvature-aware time allocation, automatically slowing at corners before a final pass enforces dynamic constraints.The curvature parameters are k_curv = 8.0 and p = 1.5.
  • C. TRAJECTORY PLANNING AND CONTROLLER: A feedforward PD controller tracks both vehicles’ generated trajectories by converting state errors into velocity commands for their inner-loop controllers.The controller serves as the outer loop for both the UGV and UAV.
  • V. SIMULATION EXPERIMENTS: High-fidelity Gazebo simulations replicated 13 m-diameter water-conveyance tunnel sections with varying curvatures and sluice gates for parameter tuning and stress testing.The environment was designed to avoid risking physical assets.
  • A. Performance Evaluation: FLISP was assessed for nominal tracking, dynamic obstacle avoidance, and runtime efficiency using representative visualizations, 100 independent trials per scenario, and quantitative runtime analyses.Paths are shown in Figs. 15–17, metrics in Table 1, and runtime distributions in Fig. 18.
  • 1) Nominal Planning Robustness: 6% of straight-section trials showed deviations, while hierarchical fitting maintained continuous floor alignment in curved sections despite sparse point clouds.UGV path stability also increased with point-cloud density.

2) Obstacle Avoidance Capabilities • … 1) Environmental Robustness and Sensor Stability

FLISP combines smooth obstacle avoidance with sub-10 ms simulated planning across normal, obstacle, and sluice-gate scenarios. Field deployment over 1.0 km exposed vibration-induced path truncation, but frequent replanning kept the effect operationally benign.

  • 2) Obstacle Avoidance Capabilities •: FLISP generated smooth UGV evasion paths under single and multiple personnel obstacles while respecting slope and path-length constraints; UAV altitude separation required minimal deviation.The planner also demonstrated robust convergence in curved, sparse-data environments.
  • 3) Runtime Efficiency Analysis: 6.2-8.8 ms obstacle-scene computation and approximately 1.5 ms sluice-gate UAV latency kept all executions within the 10 ms safety threshold.Standard-tunnel UAV planning remained generally below 1 ms; sluice-gate latency increased because of 3D obstacle segmentation and feature extraction for dive-and-recovery maneuvers.
  • B. Quantitative Metrics: Curved segments produced longer paths and more waypoints, while the UAV consistently used fewer waypoints and shorter paths than the UGV through omnidirectional, altitude-enabled corner-cutting.These metrics were summarized over 100 simulated trials.
  • VI. FIELD EXPERIMENTS: The UGV traversed 1.0 km in the functional tunnel; the final 200 m was excluded because deep silt exceeded chassis mechanical clearance, not algorithmic limits.Strict safety protocols were enforced in the confined terrain.
  • A. Experimental Platform and Environment: The experimental platform paired a self-developed UGV carrying an Intel i7-12700, GTX 4060-8G, 32 GB RAM, and Ouster OS1-64 LiDAR with a DJI Mavic 3T UAV.Additional Livox, RGB, and inspection cameras supported localization or inspection tasks.
  • A. Experimental Platform and Environment: Eight representative field scenarios stressed specific geometric features and mirrored the simulation benchmarks, excluding the continuous curve; a feedforward controller executed the generated paths.The controller converted polynomial geometry into UGV motor commands and UAV velocity setpoints transmitted via MAVLink.
  • B. Performance Analysis and Results: Field evaluation analyzed environmental robustness, dynamic obstacle avoidance, and quantitative consistency with simulation after deployment in a functional hydropower tunnel.The study therefore connected operational disturbances with the simulated performance dimensions.
  • 1) Environmental Robustness and Sensor Stability: Vibration-induced LiDAR step loss truncated distal paths by typically 2–3 m, producing 47–48 m paths instead of the 50 m horizon.Frequent replanning and the receding-horizon strategy made this marginal truncation operationally benign.

2) Robustness Validation via Handheld Obstacle Scenarios … A. Experimental Setup and Baselines

FLISP was stress-tested with handheld LiDAR jitter, showing stable path validity, lookahead, and low latency across field and simulation conditions. Comparative benchmarking used identical sensor inputs, tuned map-based baselines, and local sliding-window processing for fair evaluation.

  • 2) Robustness Validation via Handheld Obstacle Scenarios: Under severe handheld LiDAR jitter, FLISP retained total path length within 50 ± 5 m and sufficient lookahead despite 45–48 m path reductions in Scenarios 4 and 6.The reductions were attributed to jitter-induced step loss combined with obstacle occlusion.
  • 3) Runtime Efficiency and Simulation Consistency: Standard planning times remained under 10 ms for UGVs and 3 ms for UAVs, while obstacles increased UAV latency to 3–4 ms but left UGV latency invariant at ≈6 ms.UGV stability was attributed to analytically resolved path structure and one-dimensional Firefly optimization, which reduces clutter sensitivity.
  • 3) Runtime Efficiency and Simulation Consistency: Real-world execution was slightly slower and more variable than simulation, but path topology remained similar and lookahead consistently exceeded 45 m for safe high-speed tracking.Step loss reduced realized path lengths without eliminating the required lookahead.
  • VII. COMPARATIVE EXPERIMENTS AND DISCUSSION: The comparative study used a standardized offline benchmark on the 1.2 km tunnel dataset with identical sensor inputs to eliminate run-to-run physical variance.The evaluation compared FLISP with map-based paradigms under controlled conditions.
  • A. Experimental Setup and Baselines: Baselines were LIO-SAM with Informed RRT* [44] and Fast-LIO2 with Grid A* [46], alongside FLISP’s hierarchical polynomial-fitting planner.Informed RRT* was selected for its ellipsoidal heuristic and theoretical convergence acceleration in confined tunnels.
  • A. Experimental Setup and Baselines: Both map-based baselines used dynamically aligned local sliding windows from global SLAM maps, preserving sensor coverage in curved sections while bounding computational complexity.Windows were aligned with vehicle odometry from LIO-SAM or Fast-LIO2, respectively.
  • A. Experimental Setup and Baselines: Evaluation quantified smoothness, latency, and tortuosity, with latency defined as the maximum UGV or UAV cycle time and tortuosity as path length divided by horizon length.Tortuosity near 1 indicates an efficient path, and metrics were calculated within segments with valid localization to decouple planning from SLAM drift.
  • A. Experimental Setup and Baselines: To reduce SLAM-front-end stochasticity, baseline results averaged five independent trials initialized at t = 40s after IMU bias convergence.The UGV remained stationary at the tunnel entrance gate during initialization.

B. Baseline Parameter Sensitivity Analysis … 2) Root Cause and Dynamic Analysis

Baseline calibration identifies 25 m and 10.0 rad for Informed RRT* and 30 m and 0.1 m for Grid A* as operating points. In deep tunnels, mapping degradation creates distinct planning artifacts and dynamic risks, whereas FLISP preserves synchronized, low-latency collaboration.

  • 1) Method II (Informed RRT*): A 25 m horizon and 10.0 rad threshold calibrate Informed RRT* for 100% success while measuring latency at the first valid path.Success rates plummet beyond 30 m because the sampling space expands exponentially; the threshold equals the statistical upper bound µ + σ.
  • 2) Method III (Grid A*): Grid A* selects a 30 m horizon and 0.1 m resolution because longer horizons lose wall curvature through occlusion, while finer or coarser grids cause latency or blockage.A 0.05 m grid incurs approximately 369 ms latency, whereas resolutions above 0.15 m merge the floor with walls through obstacle dilation.
  • 1) Phenomenology: Mapping and Path Evolution: In deep, featureless tunnel curves, mapping diverges after the entrance, with LIO-SAM showing stochastic drift, map overlap, vertical jitter, and occasional pose decoupling.The observed LIO-SAM failure mode can render the map unusable and navigation infeasible.
  • 1) Phenomenology: Mapping and Path Evolution: Fast-LIO2 maintains localization longer than LIO-SAM but eventually develops overlapping map artifacts through longitudinal stagnation and estimator slip.The two methods therefore degrade differently despite converging to a similar overlapping artifact.
  • 1) Phenomenology: Mapping and Path Evolution: Informed RRT* produces tortuous, nonsmooth paths, whereas A* approximates the centerline but retains sawtooth discretization artifacts under yaw deviation or curvature.The irregularity of Informed RRT* is more pronounced under the looser constraints used for the UAV.
  • 2) Root Cause and Dynamic Analysis: The gate acts as a geometric anchor: when curvature occludes it, the system loses longitudinal and yaw constraints, implicating feature loss as the root mapping failure.A handheld 128-beam LiDAR control experiment was used to distinguish feature loss from vibration.
  • 2) Root Cause and Dynamic Analysis: Path artifacts destabilize the UGV through erratic steering, wall climbing, roll and vibration for Informed RRT*, and control oscillation for Grid A*.These effects compromise data quality and increase energy consumption through redundant maneuvering and gravitational resistance.
  • 2) Root Cause and Dynamic Analysis: FLISP operates at 7 ms and derives the UAV path from the UGV’s 50 m baseline to enforce spatial synchronization and reduce control phase lag.This contrasts with optimization-based planners averaging approximately 100 ms per cycle.

D. Quantitative Benchmark and Efficiency Analysis · VIII. CONCLUSION

Across five-trial benchmarks in a 1.2 km operational tunnel, FLISP achieved reliable, low-latency, resource-efficient planning while map-based baselines incurred failures, discontinuous paths, and substantial computational overhead. The conclusion presents FLISP as a lightweight mapless framework for hazardous infrastructure inspection, while identifying extensions to junctions and safety-critical low-level control.

  • D. Quantitative Benchmark and Efficiency Analysis: Method III attains lower path tortuosity (τ = 1.02) than FLISP (τ = 1.03), but FLISP has far smoother paths (0.016 rad versus 6.068 rad).The grid-based baseline’s rigid piecewise-linear paths create angular zig-zags, whereas FLISP prioritizes high-order continuity for smooth control inputs.
  • D. Quantitative Benchmark and Efficiency Analysis: 7.05 ms system latency gives FLISP a ∼7× speedup over Method III at 49.93 ms and a three-orders-of-magnitude advantage over Method II.The comparison attributes FLISP’s speed to direct geometric fitting on sparse point clouds rather than dense map maintenance.
  • D. Quantitative Benchmark and Efficiency Analysis: Over 80% of Method III’s computational budget is consumed by grid rasterization and C-space generation, not A* search itself.FLISP bypasses these map-maintenance stages, decoupling planning speed from environmental resolution requirements.
  • D. Quantitative Benchmark and Efficiency Analysis: FLISP uses 1.37% CPU load and 39.7 MB RAM, while Method III consumes 1037.3 MB and map-based planning can spend approximately 6 seconds per cycle.This footprint supports concurrent onboard perception and control on SWaP-constrained platforms.
  • D. Quantitative Benchmark and Efficiency Analysis: FLISP’s high-frequency mapless replanning mitigates momentary wheel-encoder odometry failure caused by mud, water, or loss of ground contact.It implicitly resets state relative to the latest valid LiDAR frame, addressing physical limitations without relying on a persistent map.
  • VIII. CONCLUSION: FLISP is presented as a lightweight hierarchical-polynomial, mapless framework that converts one LiDAR-IMU stream into safe synchronized UGV-UAV trajectories for degenerate infrastructure.Its feasibility was validated through comprehensive benchmarks and deployment in a 1.2 km operational tunnel.
  • VIII. CONCLUSION: Future work will add topological decision-making for complex junctions and integrate FLISP with the CBF consensus tracking controller developed in parallel work.The proposed integration is intended to combine FLISP’s reference generation with safety-critical low-level execution on the physical platform.

APPENDIX: EMPIRICAL ANALYSIS OF LOCAL PLANNER … D. Computational Efficiency

Across constrained 3D tunnels, FLISP avoids the kinematic, occlusion, topological, and computational failures observed in TEB and 2D local planners. Its geometric dimensionality reduction evaluates a 50m horizon in ∼7ms, enabling > 100Hz planning.

  • APPENDIX: EMPIRICAL ANALYSIS OF LOCAL PLANNER: The appendix benchmarks TEB against FLISP with identical LiDAR streams and safety parameters to assess kinematics, topological resilience, and computational efficiency.The evaluation targets constrained 3D tunnel environments using 0.1m resolution and 0.7m inflation.
  • A. The Heterogeneous Lookahead Paradox: Short 5m horizons produce jittery UGV paths and jagged UAV segments because heterogeneous platforms require incompatible local behaviors.UGV jitter creates high-frequency steering inputs, whereas the UAV repeatedly generates reactive straight-line segments.
  • B. The Local Occlusion Trap (Roof-Shadow Blind Spot): SLAM-based mapping can fill the roof-shadow void and avoid this specific straight-line artifact before eventual global-map collapse in deep curves.This benefit does not prevent the broader localization failure described for deep curved tunnels.
  • B. The Local Occlusion Trap (Roof-Shadow Blind Spot): TEB’s roof-shadow blind spot pulls trajectories through a perceived cost-free void, driving the UGV toward the outer wall in curved tunnels.Figure 30 shows this artifact regardless of the selected horizon; 5m horizons also cause oscillatory UAV flight and UGV control jitter.
  • B. The Local Occlusion Trap (Roof-Shadow Blind Spot): With 10m horizons, TEB produces dynamically infeasible mid-path kinks as soft constraints conflict, while curved-tunnel straight segments create rollover hazards.The straight-line artifact is benign in straight sections but unsafe when executed through curved pipes.
  • C. Topological Deadlock under Roll Perturbation: 2D costmap projection can turn a UGV roll onto a curved sidewall into an unrecoverable pseudo-collision deadlock, whereas FLISP generates a valid 3D recovery path.FLISP uses native 3D geometry and the IMU gravity vector to recover through the valid corridor shown in Figure 31.
  • D. Computational Efficiency: ∼7ms evaluation of a 50m horizon lets FLISP provide > 100Hz real-time performance, while TEB computation exceeds 100ms with high variance.The comparison used 300 planning cycles on identical onboard computing platforms; Table 5 summarizes per-update processing time.
Loading 2606.25393v2…