PaperlayerView exact revision on arXiv

Source-linked AI summary

Local Path Planning and Obstacle Avoidance for an Omnicopter Platform

Mikolaj Helinski, Spilios Theodoulis, Mahmoud Hamandi, Abdullah Mohamed Ali, Anthony Tzes, Marija Popovic

arXiv:2609.03630v1cs.RO

TL;DR

Cluttered-environment UAV navigation requires a local planner that can react at control rates despite the computational burden of omnidirectional 6D motion. This paper extends DWA with efficient mapping, geometry approximation, adaptive sampling, and context-aware weighting, achieving real-time tracking and obstacle avoidance while exposing limitations in constrained geometries.

  • Problem

    Global planners such as RRT* are not directly deployable at control rates, while unmapped or dynamic obstacles require reactive local avoidance for omnidirectional UAVs.

  • Method

    The paper extends the Dynamic Window Approach to six degrees of freedom using voxelised local maps, sphere-based vehicle approximation, adaptive velocity sampling, and Agile Mode weight switching.

  • Results

    The planner operates within a 5 Hz replanning cycle, tracks waypoint-dense paths with under 0.1 m average cross-track error and approximately 13° average orientation error, and achieves 41.4% success for a centred unknown obstacle.

  • Takeaways & Limitations

    Adaptive weighting enables diversion from global paths around unknown obstacles, while the planner maintains real-time operation and accurate waypoint-dense tracking.

  • Takeaways & Limitations

    Performance degrades in symmetric or highly constrained configurations because short-horizon reactive planning can become trapped in local minima.

Abstract

from arXiv · show

Autonomous unmanned aerial vehicles (UAVs) increasingly operate in cluttered environments where global planners such as RRT* are not directly deployable at control rates. This paper presents a real-time local planning and obstacle avoidance module for an omnidirectional multirotor (omnicopter) by extending the Dynamic Window Approach to six degrees of freedom (6D-DWA). Our method achieves real-time feasibility through (i) local-map voxelisation, (ii) a compact sphere-based approximation of the vehicle geometry, and (iii) adaptive velocity sampling in the 6D search space. To improve reactivity to unknown obstacles, we introduce a context-aware "Agile Mode" that adjusts scoring weights online to trade-off between goal progress, clearance, and heading/facing constraints during evasive manoeuvres. We evaluate our approach in simulation across computational stress tests, dense-waypoint path tracking, and static/unknown obstacle scenarios. Our planner runs consistently within a 0.2s control loop, tracks waypoint-dense global paths with < 0.1m average cross-track error and 13deg average heading error, and avoids collisions in static environments. For unknown obstacle avoidance, Agile Mode achieves 79.3% success for an off-centre obstacle and 41.4% for a centred obstacle, highlighting both the effectiveness of adaptive weighting and remaining limitations in highly constrained geometries.

I. INTRODUCTION

The paper develops a collision-aware local planner for omnidirectional multirotors because global planning and reactive avoidance must operate within real-time constraints in cluttered environments.

  • Motivation: RRT* global waypoints are unsuitable for online local replanning at control rates and may cause collisions with unmapped or dynamic obstacles.The paper also notes possible wall-induced aerodynamic suction near surfaces.
  • Approach: The proposed 6D-DWA samples feasible linear and angular velocities, rejects collision-inducing commands, and ranks the remainder with a weighted objective.The method extends DWA for omnidirectional multirotors within a broader autonomy stack.
  • Contributions: Voxelised local maps, compact vehicle-geometry approximation, and adaptive sampling enable operation within a 5 Hz replanning cycle.The paper evaluates runtime, waypoint-dense path tracking, and static and previously unmapped obstacle avoidance in simulation.
  • Motivation: Omnidirectional platforms require local planning across a six-dimensional velocity space, increasing computational demands relative to planar systems.The decision variables comprise 3D linear and angular velocities.

A. 6D Velocity Sampling

The planner makes 6D velocity search tractable by concentrating samples where promising commands are likely and accelerating collision checks with simplified geometry and voxelised maps.

  • Adaptive sampling: The 6D-DWA search space grows exponentially, so uniform sampling would increase computational load as every velocity is simulated and scored.The planner therefore limits sampling density to preserve real-time performance.
  • Adaptive sampling: Adaptive Sampling divides samples among Exploration, Focused-Search around the previous best command, and Boundary-Search at Dynamic Window limits.The three ratios satisfy er + fr + br = 1.0.
  • Collision checking: The vehicle is approximated by ten spheres with inflated radii, enabling orientation-invariant distance-based collision checks despite complex geometry.The inflation margin accounts for geometric approximation and tracking errors.
  • Collision checking: Static meshes are densified and voxelised to reduce obstacle points while preserving bounded surface resolution for collision checking.Unknown depth-camera point clouds are also voxelised and processed with KD-Trees.
  • Collision checking: A trajectory is discarded when any sampled sphere position induces a collision at any time step.The path is represented by sampled positions and corresponding sphere locations.

C. Trajectory Scoring

Trajectory scoring combines goal progress, path adherence, orientation, camera lookahead, and context-dependent obstacle terms to select among collision-free candidates.

  • Scoring framework: The planner selects the minimum weighted cost J after generating tentative trajectories.The objective combines the scoring terms described in this section.
  • Standard costs: The goal cost penalises Euclidean distance between the predicted endpoint and a local goal selected along the global path.The local goal uses a fixed lookahead offset from the current vehicle position.
  • Standard costs: The path cost measures minimum cross-track distance between predicted positions and the discretised global path.Predicted positions are projected onto path segments to compute perpendicular distance.
  • Standard costs: The heading cost penalises orientation error between the simulated pose and the local goal pose using quaternion comparison.The absolute quaternion dot product accounts for the double-cover property and shortest angular distance.
  • Sensing constraint: A lookahead cost encourages the front-facing camera to align with the path direction, which is important for sensing with a single depth camera.Perfect alignment occurs when the camera axis and lookahead direction coincide.
  • Agile Mode: Unknown-obstacle detection switches from Standard Mode to Agile Mode, adding clearance and obstacle-facing costs and prioritising clearance over strict path following.Clearance sums inverse-square distances to nearby unknown obstacle points, while facing uses the nearest obstacle direction.

D. Moving Obstacle Evasion

The moving-obstacle subsystem separates rapid evasion from 6D-DWA, using predicted collision risk and candidate manoeuvres while retaining static-environment checks.

  • Moving Obstacle Evasion: Fast-moving obstacles exceed the 6D-DWA pipeline’s reactive bandwidth because it relies on sampling-based planning.A dedicated high-frequency subsystem handles clustering, tracking, and state estimation, while threat messages are artificially generated for validation.
  • Moving Obstacle Evasion: The evasion pipeline reuses the planner’s environment representation and trajectory propagation while generating fast manoeuvres under static-environment constraints.After evasion, the system returns to 6D-DWA for nominal path following.
  • Moving Obstacle Evasion: Evasion is triggered using predicted closest-approach time that accounts for communication latency and avoids premature responses through a threshold.Collision risk is declared when predicted miss distance is below the combined bounding-sphere and obstacle radii.
  • Moving Obstacle Evasion: Candidate evasion directions are generated by rotating an initial direction around the obstacle’s velocity direction to improve robustness.The initial direction is orthogonal to the relative motion plane.

III. EXPERIMENTAL RESULTS

The experimental evaluation examines computational scaling, including how obstacle density and sample count affect real-time feasibility.

  • Experimental Results: The broader experiments assess runtime, adaptive sampling, tracking, static and unknown obstacle avoidance, and dynamic-threat evasion.These evaluations were conducted in Gazebo to assess multiple aspects of the proposed framework.
  • Experimental Results: Average computation time is evaluated against the number of obstacles using linear regression lines.The figure caption states that increased sample counts cause the real-time threshold to be exceeded at lower obstacle densities.

A. Computational Performance

Computational performance is governed by the number of sampled velocity commands and local-map obstacle points under a 0.2 s loop-time budget.

  • Computational Performance: 0.2 s is the maximum allowable loop time for the planner’s targeted 5 Hz replanning rate.Runtime is measured while varying sampled commands S and voxelised obstacle points No.
  • Computational Performance: For S = 20,000, average loop time exceeds 0.2 s when the local map contains more than approximately 54 obstacle points.The fitted runtime model associates this threshold with deteriorated tracking error.
  • Computational Performance: Increasing sample count improves planning quality but reduces allowable obstacle density under the fixed runtime budget.Larger local-map radii or smaller voxel sizes increase No and therefore runtime, so sampling should increase only while tloop remains below tmax.

B. Velocity Sampling Evaluation

Adaptive sampling improves velocity-search performance over low-sample random sampling, but insufficient exploration can degrade results in the 6D search space.

  • Velocity Sampling Evaluation: Adaptive sampling configurations outperform purely random sampling with 1,000 samples on both cross-track and orientation errors.Random sampling achieves lower errors when given more samples, while adaptive sampling improves performance for a fixed sample count.
  • Velocity Sampling Evaluation: er = 0.1, br = 0, and σ = 0.1 produced diminished performance because 90% of samples clustered around the previous best solution.The resulting concentration limited exploration of alternative velocities.
  • Velocity Sampling Evaluation: er = 0.5 and σ = 0.1 performed on-par with or better than random sampling using 5,000 samples.The selected configuration for further testing was er = 0.5, fr = 0.25, σ = 0.1, and br = 0.25.

C. 6D-DWA Weight Tuning

The 6D-DWA’s scoring weights were tuned against desired path-following behaviour, then evaluated across dense waypoint tracking and static obstacle scenarios. Results show that weight choices govern path accuracy, obstacle detouring, and traversal through orientation-constrained gaps, while controller tracking error limits safety margins.

  • Weight tuning: The grid search varied core 6D-DWA scoring weights, with each metric’s influence determined by its percentage of the total weight rather than absolute value.The trial-grading function used cross-track error, orientation heading error, covered distance, and lookahead error.
  • Weight tuning: Approximately 37% relative average wgoal and 31% relative average whead emerged across the superior runs, while lookahead-heavy grading favoured higher wlook.Higher wlook kept the camera aligned with the path, supporting environmental awareness.
  • Dense-waypoint path following: The baseline weights were set to 40, 30, 20, and 10 for wgoal, whead, wpath, and wlook, respectively, after tuning for desired behaviour.With these settings, the omnicopter maintained under 0.1 m average distance to the waypoint-dense path and approximately 13° average orientation error.
  • Static obstacle avoidance: The planner avoided commanding trajectories intersecting a wall, relying on voxelised environment representation and a sphere-based vehicle approximation for collision checking.The wall was represented in the STL map, while Gazebo did not assign it physical collision properties.
  • Static obstacle avoidance: For an obstructed global path, the planner temporarily relaxed strict path adherence, generated a safe-clearance detour, and rejoined the path after clearing the obstacle.This behaviour preserved long-term global progress while rejecting an infeasible path segment.
  • Orientation-constrained traversal: All configurations traversed the 45° gap, but baseline whead = 30 failed at the 90° gap; reducing whead enabled passage by allowing larger orientation deviations.The result shows that orientation weighting directly affects geometric feasibility in constrained 6D motion.
  • Limitations: Occasional sphere-centre proximity below the commanded safety margin was attributed to controller overshoot and tracking error, motivating radius inflation in collision checking.The observation concerns gap traversal and identifies a practical safety-margin limitation rather than a collision-checking formulation defect.

E. Agile Mode Weight Tuning

Agile Mode switches scoring weights when unknown obstacles are detected, prioritising clearance while allowing deviation from the global path. Success depended on balancing clearance, heading, and forward progress, with symmetric centred obstacles exposing local-minima limitations.

  • Agile Mode activation: Agile Mode switches from Standard Mode weights after unknown-obstacle detection, prioritising clearance over strict global-path adherence.In Agile Mode, wpath was set to zero to permit unrestricted deviation during avoidance.
  • Off-centre obstacle: 79.3% success was achieved for the off-centre obstacle with wclear = 1.0 and whead = 0.0.The baseline wclear = 1.0 configuration achieved 72.2%, while combining it with whead = 0.0 increased success to 79.3%.
  • Off-centre obstacle: Higher clearance weights increased obstacle separation but reduced forward progress, reaching 0% success at wclear = 5.0.Excessive repulsion caused stalling or timeout, demonstrating the need to balance clearance with goal-seeking behaviour.
  • Centred obstacle: 41.4% success was achieved for the centred obstacle with wclear = 1.0 and whead = 0.0, while larger clearance weights consistently failed.The symmetric configuration required committing to a lateral escape direction without strong directional bias.
  • Centred obstacle: Successful centred-obstacle trajectories predominantly diverted toward positive x coordinates, a bias attributed to the initial vehicle orientation.High repulsive forces and limited exploration capability led to local entrapment in the symmetric configuration.
  • Facing-weight effect: The facing weight did not significantly affect collision-avoidance success, but higher values delayed re-alignment with the global path.A moderate wface = 10 maintained camera alignment, whereas lower values enabled faster re-alignment after traversal.

F. Evasion Evaluation

The fast evasion pipeline handles threats that cannot wait for a standard 6D-DWA replanning cycle by predicting closest approach and validating candidate directions against static geometry. It identified collision-free manoeuvres and safely resumed global path tracking in the evaluated trials.

  • Threat evaluation: The evasion pipeline computes time to closest point of approach and minimum predicted separation from threat position and velocity estimates.This fast-reaction pipeline addresses dynamic collision threats outside the standard 6D-DWA replanning cycle.
  • Static-obstacle evasion: When the first evasion vector was blocked by static geometry, rotated candidate vectors were evaluated until a collision-free direction was identified.All trials in the static-obstacle evaluation executed a safe manoeuvre.
  • Pipeline switching: All global path-following trials successfully avoided impending collisions before control returned to 6D-DWA and global path tracking resumed.The state machine prioritised reactive safety behaviour over nominal path following during the threat.

IV. CONCLUSION

The paper presents a real-time 6D-DWA planner for omnidirectional multirotors that combines efficient sampling, geometry approximation, and voxelised local mapping. Simulations show static and unknown-obstacle avoidance with adaptive weighting, while constrained or symmetric settings remain challenging.

  • Conclusion: The planner operates within a 5 Hz replanning cycle while tracking waypoint-dense global paths using adaptive velocity sampling, sphere-based geometry approximation, and voxelised local mapping.These design choices support real-time operation for the omnidirectional multirotor platform.
  • Conclusion: Simulation results demonstrate reliable static-obstacle avoidance, unknown-obstacle diversion through context-aware weight adaptation, and fast-pipeline integration for dynamic threats.The experiments also show sensitivity to weight selection and local-minima limitations in constrained or symmetric environments.
  • Future work: Future work targets robustness in multiobstacle dynamic scenarios, uncertainty-aware safety margins, and validation on physical hardware.
Loading 2609.03630v1…