# FlightBench: A Comprehensive Benchmark of Spatial Planning Methods for Quadrotors

**Authors:** Shu-Ang Yu, Chao Yu, Feng Gao, Yi Wu, Yu Wang  
**arXiv:** [2406.05687](https://arxiv.org/abs/2406.05687)  
**Date:** 2024-06-09  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2406.05687v1/

## Key points

- Spatial planning in cluttered environments is crucial for mobile systems, especially agile quadrotors.
- FlightBench is the first comprehensive open-source benchmark designed specifically for 3D spatial planning on quadrotors.
- FlightBench allows comparison between optimization-based and learning-based methods, highlighting advantages of learning-based methods for high-speed flight and real-time planning.
- Learning-based methods show significant advantages but need improvements in trajectory smoothness and efficiency, particularly in complex conditions like navigating large corners or dealing with view occlusion.
- Latency randomization is proposed as an effective strategy to enhance performance of learning-based methods affected by pipeline latency.
- Future plans include expanding FlightBench to cover broader domains such as multi-agent planning and dynamic obstacle avoidance.

## Summary

, , , , 

Spatial planning in cluttered environments is crucial for mobile systems, particularly agile quadrotors. However, existing methods, whether optimization-based or learning-based, often focus solely on success rates in specific environments and lack a unified platform with tasks of varying difficulty. To address this gap, the authors introduce FlightBench, the first comprehensive open-source benchmark designed specifically for 3D spatial planning on quadrotors. <kf1>FlightBench:</kf1> The First Comprehensive Benchmark for Quadrotor Spatial Planning

FlightBench allows for a comparison between classical optimization-based methods and emerging learning-based approaches. It includes various representative planners and supporting modules, along with three task difficulty metrics and a thorough set of performance metrics. Through extensive experiments, the study demonstrates the significant advantages of learning-based methods for high-speed flight and real-time planning. However, it also highlights the need for improvements in trajectory smoothness and efficiency, particularly in complex conditions such as navigating large corners or dealing with view occlusion. <kf2>Advancing Learning-Based Planners:</kf2> Insights from FlightBench Experiments

The analysis further reveals that learning-based methods are particularly affected by pipeline latency. To enhance performance in real-world deployments, the authors propose latency randomization as an effective strategy. Additionally, they plan to expand FlightBench to cover broader domains including multi-agent planning and dynamic obstacle avoidance. <kf3>Critical Insights into Spatial Planning:</kf3> Lessons from FlightBench Experiments

In conclusion,<kf4>FlightBench:</kf4> provides critical insights into the strengths and weaknesses of both learning-based and optimization-based planners for quadrotor spatial planning. By offering challenging testbeds and comprehensive evaluation metrics, this benchmark aims to advance future learning-based planners in addressing the complexities of spatial planning in cluttered environments effectively.