Traditional passive point cloud acquisition systems,such as lidars or stereo cameras, can be impractical in real-lifeand industrial use cases. Firstly, some extreme environments maypreclude the use of these sensors. Secondly, they capture infor-mation from the entire scene instead of focusing on areas relevantto the end task, such as object recognition and localization. Incontrast, we propose to train a Reinforcement Learning (RL)agent with dual objectives: i) control a robot equipped with atactile (or laser) sensor to iteratively collect a few relevant pointsfrom the scene, and ii) recognize and localize objects from thesparse point cloud which has been collected. The iterative pointsampling strategy, referred to as an active sampling strategy,is jointly trained with the classifier and the pose estimator toensure efficient exploration that focuses on areas relevant to therecognition task. To achive these two objectives, we introducethree RL reward terms: classification, exploration, and poseestimation rewards. These rewards serve the purpose of offeringguidance and supervision in their respective domain, allowing usto delve into their individual impacts and contributions. We com-pare the proposed framework to both active sampling strategiesand passive hard-coded sampling strategies coupled with state-of-the-art point cloud classifiers. Furthermore, we evaluate ourframework in realistic scenarios, considering realistic and similarobjects, as well as accounting for uncertainty in the object’sposition in the workspace.