Strawberry harvesting is hard to automate because the pedicel is only 1.3–3 mm wide and similar in color to surrounding foliage. The fruit is easily bruised, so robots must detach it cleanly at the abscission zone rather than grasp the berry itself. The authors propose an imitation learning pipeline: a human teleoperates a low-cost arm to collect demonstrations, and a transformer-based policy learns to unify perception, motion planning, and contact manipulation.

Using an Action Chunking with Transformers (ACT) policy with a ResNet-50 visual backbone, the system converged from a loss of 2.84 to 0.091 over 100,000 gradient steps on 130 demonstrations. Deployed on a Jetson AGX Orin, the learned policy ran at 15 Hz. The authors note that classical approaches struggle to combine vision, motion planning, and contact handling in occluded canopy environments.

The work was accepted at an IROS 2026 workshop on agricultural robotics. The authors have released the dataset, trained weights, code, and deployment instructions, aiming to provide a reproducible baseline for the community. Since there is only one source, no independent comparison is available; the paper's claims are as reported by the authors.