In simulations, a proposed routing system for multi-hop UAV networks reached about 96% cumulative packet arrival by the deadline. Equal-Split Per-Hop Multipath reached about 90%, Capacity-Aware Per-Hop Multipath about 88%, and I-AOMDV-Guided Per-Hop Multipath about 75%. The proposed system reached nearly 98% after the deadline.
The study's main measures were the on-time packet delivery ratio and the overall packet loss ratio. The first asks how much data arrived within its required time; the second tracks how much traffic failed to arrive.
A routing system that shares traffic
MAPPO-DM is a proposed per-hop multipath method for multi-hop UAV networks. Each UAV can split traffic across multiple next hops, allowing traffic to be divided among available routes. The routing problem is framed as a Dec-POMDP, a formal model for several decision-makers operating with partial information, and MAPPO-DM is used to solve it. Its stated targets are on-time delivery and packet loss.
The decentralized Actor combines Transformer and GRU components. It generates traffic-splitting actions through a Dirichlet model, which represents how the traffic share is divided among available next hops. A centralized graph-attention Critic evaluates the wider network during training.
The reference environment used 35 UAV nodes. Simulated time advanced in 0.5-second slots, each episode lasted 60 seconds, and routing decisions allowed up to six candidate neighbors.
Each simulated task contained 1.0 to 2.0 megabytes of data, divided into 1,500-byte packets, with routing deadlines from 8 to 14 seconds.
The evaluation included four representative routing baselines, covering per-hop multipath and heuristic single-path approaches. The named multipath comparisons were Equal-Split, Capacity-Aware and I-AOMDV-Guided routing.
The deadline is where the gap matters
When individual components were removed, the complete system was not highest on every measure. On-time delivery was 96.70% without the adaptive μ-κ Dirichlet concentration, 96.40% for full MAPPO-DM, 96.25% without the Transformer and 95.91% without the Graph Critic. The fixed-concentration version therefore had the highest reported on-time figure in this comparison.
Packet loss showed a different ranking. MAPPO-DM and the no-Graph-Critic variant both recorded 0.48%, compared with 0.84% without the Transformer and 1.78% without the adaptive μ-κ Dirichlet concentration. In this component test, the complete system tied for the lowest reported loss.
The cumulative-arrival comparison makes timing visible alongside eventual delivery. At the deadline, MAPPO-DM stood at about 96%, versus about 90% for Equal-Split, about 88% for Capacity-Aware and about 75% for I-AOMDV-Guided. It reached nearly 98% afterward. The plotted values show the proposed method accumulating arrivals more quickly near the time limit, although the reported figures are approximate.
Changing the conditions changes the picture
Stress tests changed hotspot regions and traffic load. Every algorithm showed lower on-time delivery and higher packet loss as load rose. MAPPO-DM retained the best reported performance and degraded more slowly.
Changing network size produced less uniform results. As the network grew, MAPPO-DM, Equal-Split and Capacity-Aware generally benefited from a denser topology, while performance changed only slightly between 35 and 40 UAVs. I-AOMDV-Guided improved at first and then degraded.
Varying the candidate-forwarding limit produced a small initial improvement followed by nearly unchanged performance as the limit increased.
The authors report MAPPO-DM as the best overall scheme among those compared, with better latency guarantees and lower packet loss across the tested network conditions. That conclusion is limited to the reported simulations and baselines.
What the simulation leaves open
Each environment setting was averaged over 50 independent runs to reduce randomness. The report gives descriptive mean comparisons, but no confidence intervals or formal significance tests, so the percentage gaps should be read as reported differences rather than statistically tested ones.
The reported evidence is limited to the modeled simulation conditions and the baselines tested. It does not show how the method would perform on physical UAV hardware or outside the tested scenario ranges.
The document header identifies the work as arXiv:2608.25383v1 in cs.NI, dated 26 August 2026.
Source code for both training and testing is stated to be publicly available online.
No funder is identified in the supplied material. The acknowledgment reports use of OpenAI ChatGPT for language editing and limited code assistance, and says the authors verified the technical content and results.
Paper data and sources
Original title: Traffic-Adaptive Per-Hop Multipath Routing in Multi-Hop UAV Networks
Authors: Zhenyu Zhao, Tiankui Zhang, Xiaoxia Xu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text