A tomato-picking robot achieves 81% success rate by predicting how easy each tomato will be to harvest before attempting to pick it. Instead of applying the same strategy to every fruit, it adjusts its approach based on difficulty assessment.
Most robotic harvesting systems treat every target identically: detect, approach, grasp, pull. The failure mode is that tomatoes vary — in ripeness, stem strength, cluster density, leaf occlusion, and branch flexibility. A picking strategy optimized for easy tomatoes fails on difficult ones, and vice versa. The universal approach hits a ceiling because the task is not uniform.
The 81% success rate comes from a prediction layer that classifies each tomato's anticipated difficulty before the robot commits to a picking strategy. Easy tomatoes get a fast, direct approach. Difficult ones — occluded, tightly clustered, or with tough stems — get a modified path, different grip parameters, or are skipped entirely for later.
The insight is broader than agriculture. When a system faces a heterogeneous task population, the optimal strategy is not to find the best single approach but to classify first and specialize second. Diagnostic before treatment. Assessment before action.
The robot's improvement came not from a better gripper or a faster arm but from the decision to look before reaching. The bottleneck was judgment, not execution.