Abstract: In this paper, we study the task of searching for a query object of unknown position and pose in a scene, both given in the form of 3D point cloud data. A straightforward approach that exhaustively scans the scene is often prohibitive due to computational inefficiencies. High-quality feature representation also needs to be learned to achieve accurate recognition and localization. Aiming to address these two fundamental problems in a unified framework, we propose Adaptive Spotting, a deep reinforcement learning approach that jointly learns both the features and the efficient search path. Our network is designed to directly take raw point cloud data of the query object and the search bounding box and to sequentially determine the next pose to be searched. This network is successfully trained in an end-to-end manner by integrating a contrastive loss and a reinforcement localization reward. Evaluations on ModelNet40 and Stanford 2D-3D-S datasets demonstrate the superiority of the proposed approach over several state-of-the-art baselines.