A practical obstacle detection system for autonomous orchard vehicles

Safe robot navigation in tree fruit orchards requires that the vehicle be capable of robustly navigating between rows of trees and turning from one aisle to another; that the vehicle be dynamically stable, especially when carrying workers; and that the vehicle be able to detect obstacles on its way and adjust its speed accordingly. In this paper we address the latter, in particular the problem of detecting people and apple bins in the aisles between rows. One of our requirements is that the obstacle avoidance subsystem shouldn't add to the robot's hardware cost, so as to keep the acquisition cost to growers as low as possible. Therefore, we confine ourselves to solutions that use only the sensor suite already installed on the robot for navigation-in our case, a laser scanner, low-cost inertial measurement unit, and steering and wheel encoders. Our methodology is based on the classification and clustering of registered 3D points as obstacles. In the current implementation, obstacle avoidance takes in 3D point clouds collected in apple orchards and generates an off-line assessment of obstacle position. Tests conducted at our experimental orchard-like environment in Pittsburgh and an actual apple orchard in Washington state indicate that the method is able to detect people and bins located along the vehicle path. Stretch tests indicate that it is also capable of dealing with objects as small as 15 cm tall as long as they aren't covered by grass, and to detect people crossing the aisles at walking speed.