Efficient Sample Collection to Construct Observation Models for Contact-Based Object Pose Estimation

Abstract

It is important for a robot to accurately estimate the pose of an object in order to manipulate it. Estimation by tactile information, which is not affected by occlusion, provides more valid estimates than visual information in some cases. To realize such tactile sensing-based object pose estimation, it is necessary to construct an observation model in advance by collecting samples of contact action and observed information. This process of sample collection is generally time-consuming and its cost can be a disadvantage of the tactile sensing-based object pose estimation. To mitigate the cost, in this paper, we propose an efficient sample collection method to generate observation models for object pose estimation. Contact actions useful for pose estimation are quantitatively evaluated, and efficient sample collection is achieved by a search to find the point that minimizes the quantitative value. The proposed sample collection strategy was evaluated in simulations by assuming a specific shape of object with a soft tactile sensor. It was shown that the proposed strategy realizes more efficient sampling in comparison with random/uniform sampling methods.