REAPER: ARTICULATED OBJECT 6D POSE ESTIMATION WITH DEEP REINFORCEMENT LEARNING

Current articulated object pose estimation methods largely rely on dense prediction for all the input observed point cloud that suffers from huge computational costs and inference time. Besides, self-occlusion is also becoming a key problem that limits the pose estimation performance for those child parts. To solve these issues, we propose a Reinforcement learning based Articulation Pose EstimatoR (ReAPER), which integrates RL into deep neural network for per-part pose estimation. Specifically, we design the novel action space that involves the object's rotation and translation, as well as a reward function considering chamfer distance during pose fitting. To speed up the RL policy training, we employ imitation learning for policy initialization. Finally, we also introduce a new kinematic energy function to optimize the child parts' poses. Experimental results show that ReAPER could obtain state-of-the-art performance on articulated object pose estimation task.