Advancements in Multi-Agent Robotic Systems for 3D Reach-Avoid Games
New methods improve assignment quality in multi-agent robotic interactions.
This research explores the quality of assignments in 3D heterogeneous multi-agent reach-avoid games, highlighting a failure mode known as Geometric Sprawl. This occurs when multiple maximum-cardinality assignments lead to inefficient spatial pairings and pursuit paths.
To address this, a cardinality-first weighted sequential matching method is proposed, utilizing the Hamilton--Jacobi--Isaacs interception value as a secondary weight. The study evaluates both weighted and unweighted methods across a 35-scenario benchmark, demonstrating significant improvements in capture rates and pursuit efficiency with the weighted approach.
The findings indicate a 22% increase in mean captures and a 40% increase in mean interception height, alongside smoother pursuit trajectories. The simulator and benchmark suite are made available for further research.