A planner built around rows
An orchard-navigation planner called AGRO-Nav reported less drift from tree-row centers than the comparison planners in a field dataset, while also taking less time to answer each planning query, according to an arXiv preprint. In that evaluation, mean row-center error was 0.08 m for AGRO-Nav, versus 0.31 m for A* and 0.43 m for Theta*, and planning time was 11.52 ms versus 51.24 ms and 58.48 ms.
The paper asks whether an automatically constructed orchard row-topology graph can support global, row-centered planning without manual graph waypoints. To build it, AGRO-Nav fits tree-row lines from a SLAM point cloud, builds connections within and between rows, routes with Dijkstra, links off-graph start and goal poses with Theta* segments, and smooths the route with a cubic B-spline.
What the comparisons showed
In the simulation setup, six start–goal pairs were evaluated. Six representative rows were sampled at 11 points each, giving 66 evaluation points. At full tree density, AGRO-Nav recorded the lowest mean and maximum row-center errors among all compared methods: 0.14 m and 0.42 m.
All planners used the same 2D occupancy grid and identical conditions. Each was run ten times per start–goal pair, with path metrics averaged across pairs and planning time averaged across repetitions.
In the reported simulation settings, AGRO-Nav also had the greater minimum clearance from tree trunks when compared with reproduced RANSAC. At full density, its path came no closer than 1.93 m to a trunk, compared with 1.86 m for RANSAC. At 70% tree density, the corresponding figures were 2.03 m and 1.86 m. AGRO-Nav’s planning times in those settings were 14.24 ms and 13.48 ms.
Reducing simulated tree density to 70% did not change the reported result for AGRO-Nav: it remained the lowest-error method among A*, Theta*, and reproduced RANSAC and was described as stable under the reduction. RANSAC’s mean error increased from 0.18 m to 0.24 m when density was reduced, which the authors suggest reflects greater sensitivity to sparse tree distributions.
The RANSAC comparison carries a qualification. Its original source code was not publicly available, so the authors reproduced the baseline from the published algorithmic description and described it as representative rather than official.
The trade-off behind the numbers
The real-orchard evaluation used three start–goal pairs and 33 row-center evaluation points, with 11 points per row. Because ground-truth trajectories were unavailable, row centerlines served as the field reference, or pseudo-ground truth.
That choice is important when reading the headline error figures. The main metric measures row adherence rather than general path optimality, and it structurally favors a planner designed to center its route in the rows.
The row-centered objective also came with a slightly longer real-orchard trajectory. AGRO-Nav’s path measured 218.03 m, versus 214.09 m for A* and 212.70 m for Theta*. The authors describe the extra length as an intended trade-off for staying nearer the row center.
A narrow test, not a deployment trial
The findings should be read as evidence about static global planning in the tested settings. Only static global planning was evaluated, and the real-world evaluation was conducted at a single commercial orchard site.
The document is an arXiv preprint, version 1, dated 26 Aug 2026. No funding statement was reported in the supplied text or metadata. Within those limits, the preprint reports a planner that stayed close to row centers and answered the tested planning queries quickly in both simulation and the single orchard evaluation.
Paper data and sources
Original title: AGRO-Nav: Autonomous Graph-based Orchard Navigation
Authors: Ho Young Yun, Jaemin Yu, Duksu Kim
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text