# Multi-UAV Path Planning for Wireless Data Harvesting with Deep Reinforcement Learning

**Authors:** Harald Bayerlein, Mirco Theile, Marco Caccamo, David Gesbert  
**arXiv:** [2010.12461](https://arxiv.org/abs/2010.12461)  
**Date:** 2020-10-23  
**Summary page:** https://www.summarizepaper.com/en/paper/2010.12461v3/

## Key points

- Challenging problem of harvesting data from distributed IoT devices using multiple UAVs
- Proposed MARL approach for adapting to changes in scenario parameters without recomputations or relearning control policies
- Formulated path planning problem as Dec-POMDP and solved it through DRL
- Effective cooperation among agents enabled by proposed network architecture
- Balance between data collection goals, flight-time efficiency, and navigation constraints in movement decisions
- Control policy that generalizes over scenario parameter space for analyzing individual parameter influence on collection performance and system-level benefits
- Code availability for further exploration provided
- Contributes to addressing challenges of data harvesting from IoT devices using UAVs
- Offers insights into optimizing path planning in complex environments.

## Summary

The paper discusses the challenging problem of harvesting data from distributed Internet of Things (IoT) devices using multiple autonomous unmanned aerial vehicles (UAVs). The authors propose a multi-agent reinforcement learning (MARL) approach that can adapt to changes in the scenario parameters without the need for expensive recomputations or relearning control policies. The path planning problem is formulated as a decentralized partially observable Markov decision process (Dec-POMDP), which is solved through deep reinforcement learning (DRL). The authors demonstrate that their proposed network architecture enables effective cooperation among the agents, allowing them to divide the data collection task and make movement decisions that balance data collection goals, flight-time efficiency, and navigation constraints. By learning a control policy that generalizes over the scenario parameter space, the authors are able to analyze the influence of individual parameters on collection performance and provide insights into system-level benefits. Furthermore, this research extends previous work by providing code availability for further exploration. Overall, this paper contributes to addressing the challenges of data harvesting from IoT devices using UAVs and offers valuable insights into optimizing path planning in complex environments.