Object Perception Guided Autonomous Exploration

With the rapid progress of robotics,robots have gradually stepped into people’s daily lives and brought intelligent services to humans.Efficiently perceiving objects in unknown indoor scenes with high quality is a prerequisite for many robotic tasks.Compared with the previous indoor exploration algorithms assuming a human scanner with a hand-held sensor,researchers become more interested in automatically conducting exploration with a robot instead of a human.In order to perceive objects efficiently in the exploration of unknown indoor scenes,an object perception oriented autonomous exploration algorithm for a robot is proposed.Using deep reinforcement learning,the robot learns to use the layout rules and semantic information of the scene to obtain a more efficient and high-quality exploration strategy through interaction with the environment.Our key insight is that when humans explore an unknown indoor scene,they not only make a global decision on a long-term goal to go to,but also adjust their view on the way to abtain more information.Based on the insight,this paper optimizes the orientation of the sensor in local exploration module so that the robot can adjust the orientation of the sensor during the exploration process to obtain more object perception oriented information as humans do.In the specific implementation,this paper uses a modular framework to conquer the difficulty in the training of deep reinforcement learning,which is divided into simultaneous localization and mapping module,global exploration module,path planning module and local exploration module.Simultaneous localization and mapping module constructs a scene map based on data obtained by sensor and information obtained from the object instance segmentation module.Then the global exploration module decides a long term goal based on the scene map to guide the robot to the area to be explored.Next,the path planning module is employed to generate a collision-free trajectory for robot navigation.And the local exploration module plans the orientation of the sensor at each step based on local map around the robot and updates the map.We carry comparative experiment with some advanced algorithms in simulation environment.Compared with other exploration algorithms,in which the orientation of the sensor is always consistent with the orientation of the robot,the local exploration in our paper takes the rotation of sensor relative to the robot into consideration and make the sensor more conducive to the direction of objects to be perceived.The experimental results show that the method in this paper can improve the object perception rate,perception efficiency and object detection accuracy of the object instance segmentation network.