Researchers at Oregon State University trained a robot arm on computer-generated trees, then tested whether it could guide a pruning tool to branches in real orchards. It reached the specified position and angle in 10 of 14 trials in a commercial apple orchard without building a detailed 3D map of each tree. The field result is a step beyond the team’s earlier simulation work, though the trials tested tool positioning rather than an automated service pruning whole trees.
A camera guides the tool through branches
Winter pruning is difficult to automate because branches, posts and trellis wires crowd the route to a chosen cut point. A conventional planner first reconstructs that clutter in 3D; thin or overlapping branches can be missed or merged in the resulting map. This study instead used optical flow—the apparent movement of features between frames from a camera mounted beside the cutter—to adjust the arm as it moved.
The researchers generated synthetic apple and cherry trees and used a simulator to collect successful routes around their branches. Their training method combined those example routes with further trial-and-error learning in simulation. The controller was then transferred to a physical arm without retraining it on orchard footage. It was given a target cut point; it did not decide which branches a grower should remove.
Field success varied between two orchard designs
In the commercial V-trellis apple orchard, the controller met the study’s position and orientation thresholds in 10 of 14 attempts. It did so in 5 of 14 trials in an experimental cherry orchard whose steeper canopy departed more from the training setup. The researchers selected field targets within the arm’s reachable area and favored relatively safe configurations, so these rates do not describe every branch in an orchard.
A separate, less selectively chosen laboratory comparison used the same 15 targets for both systems. The learned controller succeeded on 7, versus 4 for a conventional planner using a reconstructed 3D point cloud. That comparison took place indoors, not against another robot in the commercial orchard. Across successful field and lab runs, a tool-positioning trajectory took 41 seconds on average.
The arm still stopped at difficult joint positions, timed out before aligning the cutter, or struck orchard structures. Its wrist camera also could not see every part of the arm. The tests show that training on synthetic trees can transfer to outdoor tool positioning, while leaving branch selection, reliable collision avoidance and practical pruning speed unresolved.
Sources
- Oregon State University research paper, arXiv:2609.24906 (submitted September 21, 2026).
- Oregon State University TreeSim source code, used to generate the orchard tree models.
- Lead image: Figure 1 from the research paper, by Abhinav Jain, Cindy Grimm and Stefan Lee, reused under CC BY 4.0; no changes.