A GPU-based solver reported the lowest total flight-time cost for UAV routes in simulated agricultural grids, while the authors reported calculation times below 320 milliseconds across the tested grid dimensions. The finding is a computational result, not a demonstration on a physical drone or in a working field.
The work is an arXiv preprint, version 1, dated 26 Aug 2026. Its authors present the method as a real-time, scalable approach for agricultural spraying, while the benchmark compares algorithms on modeled grids.
Turning a flight route into a computer puzzle
The proposed system converts grid parameters and constraints into a QUBO model, a format that reduces route planning to linked yes-or-no decisions. It then sends the model to Fixstars Amplify’s GPU-based Annealing Engine.
The formulation requires one cell to be occupied at every time step and each cell to be visited exactly once. It also uses binary decision variables, so each possible cell-and-time assignment is selected or not selected. The UAV’s starting position is fixed at x0,0 = 1.
The simulations used grid cells measuring 5.0 metres and a UAV travel speed of 5.0 metres per second. The Amplify Annealing Engine was given an optimisation time limit of 100 milliseconds.
The comparison was close on small grids
The computational benchmark covered listed grid configurations ranging from 9 to 36 nodes. It compared the GPU solver with CPU-based simulated annealing using the same QUBO formulation, as well as a genetic algorithm using crossover and mutation.
For total flight-time cost, the proposed method ranked first, followed by the genetic algorithm and then simulated annealing. The methods were described as nearly equivalent on small grids, with clearer differences appearing as the grid expanded. Exact flight-time values and run-to-run variability were not reported.
The settings gave simulated annealing 1,000 samples and a random seed of 25. The genetic algorithm used a population of 512, 10 crossover parents, a mutation probability of 0.2, an adjusted power coefficient of 1.0 and 1,000 generations.
A speed advantage with important qualifiers
Calculation time was described as stable for the proposed solver as the number of nodes increased. The comparison methods generally took longer on larger configurations. The genetic algorithm was slowest overall, while simulated annealing was fastest on the small configurations.
Across grids from 3 by 3 to 6 by 6 cells, the authors report that the proposed method identified optimal paths in less than 320 milliseconds. They attribute the reported solution-quality advantage to parallel annealing and broader exploration of the model’s energy space, which they say can help the search escape local optima.
The route planner is not yet a field system
The study uses a simplified flight-time Hamiltonian. The authors identify battery capacity, pesticide application amount and larger grids as extensions for future work, so those added physical and operational constraints are outside the reported scope.
The evidence is limited to numerical simulations and algorithmic comparisons on modeled agricultural grids. The supplied analysis reports no physical UAV flight, field deployment or agricultural outcome measurement, and it also notes that exact calculation-time values, independent-run counts and variability were not provided.
A further technical check remains: the displayed coverage and penalty sums appear to stop at n-2 even though the time-step set and mission description extend through n-1. That indexing issue should be resolved before the paths are treated as fully validated under the intended formulation.
The next tests must establish whether the reported rankings hold across repeated independent runs, larger agricultural environments and implementations that include battery and spraying constraints.
Paper data and sources
Original title: GPU-Accelerated Quantum Annealing-Inspired UAV Path Planning for Smart Agriculture
Authors: Maho Hirahara, Aohan Li
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text