[Paper Review] Comparative Analysis of Agent-Oriented Task Assignment and Path Planning Algorithms Applied to Drone Swarms

Rohith Gandhi Ganesan, Samantha Kappagoda+2 more|arXiv (Cornell University)|Jan 13, 2021

Robotic Path Planning Algorithms15 references4 citations

TL;DR

This paper proposes a decentralized, reinforcement learning-based framework for drone swarm task assignment and path planning in unknown environments, coupling policy gradient networks with Wavefront and Potential Field planners. It introduces adaptive velocity control via Huber loss to improve collision avoidance, achieving faster coverage (309 time steps) than baseline L1 loss (561 steps) in 3D simulations with 2 drones and 25 PoIs.

ABSTRACT

Autonomous drone swarms are a burgeoning technology with significant applications in the field of mapping, inspection, transportation and monitoring. To complete a task, each drone has to accomplish a sub-goal within the context of the overall task at hand and navigate through the environment by avoiding collision with obstacles and with other agents in the environment. In this work, we choose the task of optimal coverage of an environment with drone swarms where the global knowledge of the goal states and its positions are known but not of the obstacles. The drones have to choose the Points of Interest (PoI) present in the environment to visit, along with the order to be visited to ensure fast coverage. We model this task in a simulation and use an agent-oriented approach to solve the problem. We evaluate different policy networks trained with reinforcement learning algorithms based on their effectiveness, i.e. time taken to map the area and efficiency, i.e. computational requirements. We couple the task assignment with path planning in an unique way for performing collision avoidance during navigation and compare a grid-based global planning algorithm, i.e. Wavefront and a gradient-based local planning algorithm, i.e. Potential Field. We also evaluate the Potential Field planning algorithm with different cost functions, propose a method to adaptively modify the velocity of the drone when using the Huber loss function to perform collision avoidance and observe its effect on the trajectory of the drones. We demonstrate our experiments in 2D and 3D simulations.

Motivation & Objective

• To address the challenge of optimal coverage in unknown environments using drone swarms with limited global knowledge.
• To compare deep reinforcement learning architectures for decentralized task assignment in multi-drone systems.
• To integrate task assignment with local path planning for real-time obstacle avoidance without prior environmental knowledge.
• To evaluate the impact of adaptive velocity control using Huber loss on trajectory safety and coverage efficiency.
• To benchmark performance across different policy networks and path planning strategies in 2D and 3D simulations.

Proposed method

• Employed policy gradient reinforcement learning with sequence-to-sequence, attention-augmented sequence-to-sequence, Pointer Net, and Transformer networks for task assignment.
• Coupled task assignment with two path planning strategies: grid-based Wavefront for global planning and gradient-based Potential Field for local obstacle avoidance.
• Proposed a dynamic velocity scaling method using the Huber loss function, where δ is adaptively adjusted based on the intersection of field-of-view and obstacle volumes.
• Defined δ as δ = (Vol(FoV)) / (Vol(FoV) + α(Vol(FoV) ∩ Vol(Obstacle))^β) × d/2 to modulate step size during navigation.
• Used L1 loss as baseline and compared performance with Huber loss under varying α and β parameters in 3D simulations.
• Evaluated all configurations across three simulation runs, measuring time steps as a proxy for coverage time and efficiency.

Experimental results

Research questions

• RQ1 Which deep reinforcement learning architecture yields the most effective and efficient task assignment for drone swarms in unknown environments?
• RQ2 How does adaptive velocity control via Huber loss impact collision avoidance and coverage time in drone path planning?
• RQ3 What is the comparative performance of Wavefront and Potential Field planners when integrated with RL-based task assignment?
• RQ4 How do different cost functions in the Potential Field planner affect trajectory stability and safety?
• RQ5 Can a hybrid approach of decentralized task assignment and local path planning achieve faster and safer coverage than baseline methods?

Key findings

• The use of adaptive δ with α=1 and β=1 in the Huber loss function reduced time steps to 309, outperforming the L1 loss baseline of 561 time steps in 3D simulation.
• Higher α and β values (e.g., α=4, β=1) led to overly conservative behavior, increasing time steps to 469 due to excessively small velocity updates.
• The Transformer-based policy network showed strong performance, though specific comparative metrics were not detailed beyond time steps and computational load.
• The Potential Field planner with adaptive δ achieved better coverage speed and smoother trajectories than fixed-velocity or non-adaptive strategies.
• The Wavefront planner provided robust global path planning but was less effective in dynamic or dense obstacle environments compared to the gradient-based Potential Field.
• The combination of RL-based task assignment and adaptive local planning significantly reduced coverage time and improved safety without requiring global obstacle knowledge.