An analysis of sensor selection for fruit picking with suction-based grippers
Researchers identify minimal sensors to predict fruit-picking failures before they happen...
In a new paper submitted to arXiv, researchers Eva Krueger, Marcus Rosette, and Joseph R. Davidson tackle a key challenge in agricultural robotics: reliably detecting whether a fruit has been successfully picked. Their solution is a multimodal sensing suite integrated into a compliant suction-based apple gripper. The system uses a phase-dependent approach, identifying which sensors are most informative at different stages of the pick to predict failures before they occur.
Tested in a real apple orchard, the system achieved over 90% accuracy with both Random Forest and Multilayer Perceptron classifiers. The Random Forest model predicted pick and slip events within 0.09 seconds of human-annotated ground truth. This work highlights minimal sensor sets for reliable pick state classification, potentially reducing crop damage and boosting efficiency in automated harvesting.
- Random Forest and MLP classifiers detect pick success and failures with over 90% accuracy
- System predicts pick/slip events within 0.09 seconds of human annotations
- Phase-dependent sensor selection minimizes sensor sets while maintaining reliability
Why It Matters
This could reduce crop damage and improve efficiency in automated fruit harvesting, addressing a major bottleneck in agricultural robotics.