Cascaded Deep Search: Achieving High Object Localization Precision for Roughly Positioned Autonomous Mobile Manipulator

Autonomous Mobile Manipulators (AMMs) transport between locations and execute manipulation tasks such as object handling, As the manipulator resides on the mobile platform, the performance of manipulation is often limited by the position accuracy of the mobile platform. The robot can often be poorly positioned for dynamically balanced robots or those without rigid tracks. As a result, there exists a high uncertainty on the relative position between the target and AMMs parking position. This article contemplated the above scenario and proposed a cascaded visual object localization architecture to achieve vast-area search and progressive localization precision. We demonstrated a three-stage cascaded design utilizing the Convolutional Neural Network for illumination-robust localization in the final stages. Our field tests showed that given approximately 60×20 cm
²
position variations of the target object, the proposed system located the target object and performed pick-and-place with a success rate of 95%. The method was demonstrated on a two-wheeled mobile robot; however, it can also be applied to other mobile platforms that lack positioning accuracy due to rugged terrains or other reasons. The entire localization module is light and operates on a Raspberry Pi 4.