Simulation, Learning and Control Methods to Improve Robotic Vegetable Harvesting

Abstract

Agricultural robots are subject to a much harsher environment than those in the factory or lab and control strategies need to take this into account while maintaining a low cycle time. Three control strategies were tested on Vegebot, a lettuce-picking robot, in both simulation and on the real robot. Between a fast open loop that was vulnerable to environmental noise and a slow but robust visual servoing technique, a Learned Open Loop strategy was tested where the robot learned from successful picks to pick at an intermediate speed. This reduced the projected cycle time from 31s to 17.2s, a 45% reduction.