Reducing Time in Active Visual Target Search with Bayesian Optimization for Robotic Manipulation

For robotic tasks in unknown and large environments, it is important to locate targets to be manipulated by the robot in the environment. If the workspace is too large to capture at a time within the limited measurement field of visual sensors, it is necessary to explore the workspace by moving sensors. However, previous research which models almost the whole environment conduct unnecessarily exhaustive exploration for task execution. The purpose of this research is to enable robots to locate targets in a large workspace in a short time for task execution. Target exploration is formulated as a variation of a 2D black-box optimization problem. A sample-efficient exploration algorithm is proposed based on Bayesian optimization(BO) with Gaussian Process(GP). To utilize the continuity of physical motion of the camera, the algorithm for continuous search is proposed. Furthermore, to utilize prior knowledge about target positions, if available, a method of incorporating prior distribution in GP is proposed, which enables even faster exploration. Experimental results in both simulation and real environments demonstrated the efficiency of the proposed method compared to hand-crafted or naive BO-based exploration algorithms.