Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Multi-UAVs end-to-end Distributed Trajectory Generation over Point Cloud Data

Antonio Marino, Claudio Pacchierotti, Paolo Robuffo Giordano

TL;DR

This work addresses scalable, collision-free trajectory generation for teams of UAVs navigating cluttered, dynamic environments using LiDAR point-cloud data. It introduces a decentralized, end-to-end planner with two neural branches—one proposing a trajectory in MINVO CP space and another predicting collision constraints—augmented by a differentiable QP layer and MAGAT-based communication. Training leverages a privileged expert (MADER) to create robust datasets, and a point-cloud saliency map (PointBackProp variant) offers interpretability for the learned decisions. Experiments demonstrate strong performance up to 25 UAVs and 25% obstacle density, with 100% success in moderate scenarios and competitive travel times compared to baselines, along with successful physical-simulation validation. The approach offers practical impact for real-time multi-UAV coordination by combining end-to-end learning, decentralized communication, and safety guarantees, while providing a toolset for understanding perception-driven decisions via saliency maps.

Abstract

This paper introduces an end-to-end trajectory planning algorithm tailored for multi-UAV systems that generates collision-free trajectories in environments populated with both static and dynamic obstacles, leveraging point cloud data. Our approach consists of a 2-fork neural network fed with sensing and localization data, able to communicate intermediate learned features among the agents. One network branch crafts an initial collision-free trajectory estimate, while the other devises a neural collision constraint for subsequent optimization, ensuring trajectory continuity and adherence to physicalactuation limits. Extensive simulations in challenging cluttered environments, involving up to 25 robots and 25% obstacle density, show a collision avoidance success rate in the range of 100 -- 85%. Finally, we introduce a saliency map computation method acting on the point cloud data, offering qualitative insights into our methodology.

Multi-UAVs end-to-end Distributed Trajectory Generation over Point Cloud Data

TL;DR

This work addresses scalable, collision-free trajectory generation for teams of UAVs navigating cluttered, dynamic environments using LiDAR point-cloud data. It introduces a decentralized, end-to-end planner with two neural branches—one proposing a trajectory in MINVO CP space and another predicting collision constraints—augmented by a differentiable QP layer and MAGAT-based communication. Training leverages a privileged expert (MADER) to create robust datasets, and a point-cloud saliency map (PointBackProp variant) offers interpretability for the learned decisions. Experiments demonstrate strong performance up to 25 UAVs and 25% obstacle density, with 100% success in moderate scenarios and competitive travel times compared to baselines, along with successful physical-simulation validation. The approach offers practical impact for real-time multi-UAV coordination by combining end-to-end learning, decentralized communication, and safety guarantees, while providing a toolset for understanding perception-driven decisions via saliency maps.

Abstract

This paper introduces an end-to-end trajectory planning algorithm tailored for multi-UAV systems that generates collision-free trajectories in environments populated with both static and dynamic obstacles, leveraging point cloud data. Our approach consists of a 2-fork neural network fed with sensing and localization data, able to communicate intermediate learned features among the agents. One network branch crafts an initial collision-free trajectory estimate, while the other devises a neural collision constraint for subsequent optimization, ensuring trajectory continuity and adherence to physicalactuation limits. Extensive simulations in challenging cluttered environments, involving up to 25 robots and 25% obstacle density, show a collision avoidance success rate in the range of 100 -- 85%. Finally, we introduce a saliency map computation method acting on the point cloud data, offering qualitative insights into our methodology.
Paper Structure (13 sections, 7 equations, 6 figures, 2 algorithms)

This paper contains 13 sections, 7 equations, 6 figures, 2 algorithms.

Table of Contents

  1. INTRODUCTION
  2. PROBLEM STATEMENT
  3. METHOD
  4. Trajectory Representation
  5. Privileged Expert
  6. Neural Network Input Features
  7. Neural Network Structure
  8. Point Cloud Saliency Map
  9. TRAINING
  10. EXPERIMENTS
  11. Results
  12. Physical Simulation
  13. CONCLUSION

Figures (6)

  • Figure 1: neural network architecture deployed on each drone.
  • Figure 2: Experimental environments: dynamic corridor (right) and dynamic forest (left).
  • Figure 3: Success rate and safety rate increasing obstacles and agents in the range of obstacle density [$5\% - 25\%$] and number of robots between [$5-25$] for our approach, ours without GNN, ours without collision constraint, ours with Pointnet++, ours without DMoN and MADER algorithm.
  • Figure 4: Saliency map in a scenario with two drones for our approach and ours without GNN. The two drones (black and blue) sense each other and a pilar through their point cloud while moving toward the target. The saliency map is displaced by the alpha channel of the points.
  • Figure 5: Saliency map (on the left) of collision case using our approach for a trajectory starting from the drone and traversing the obstacle.
  • ...and 1 more figures

Theorems & Definitions (1)

  • Definition 3.1