View Planning for Object Pose Estimation Using Point Clouds: An Active Robot Perception Approach

Abstract

This letter considers the object pose estimation problem for robotic tasks where the robot end-effector is mounted with a vision sensor to collect point clouds from multiple views. The focus of this work is on generating and planning sensor views online based on the feedback from point cloud analysis to improve point cloud quality and quantity for pose estimation. To this end, we propose a novel active robot perception framework to generate sensor views online based on the collected data to improve the estimated pose. This framework includes the following components: evaluating the quality and quantity of the point clouds, finding a minimum number of optimal sensor views, and calculating valid robot poses corresponding to the sensor views. Tools and techniques from combinatorial optimization, mixed-integer programming, and constrained nonlinear optimization are utilized to formulate and solve the sensor view generation and robot pose problems. Extensive experiments were conducted to evaluate the proposed framework, and a representative sample of the results are provided. For baseline comparison, the proposed method is compared with two view planning methods from reconstruction, and the results show the benefits of incorporating the proposed active feedback mechanism for pose estimation. The overall effectiveness of the active perception strategy is shown through improved pose estimation accuracy.