Active Object Detection Based on PPO Learning Algorithm with Decision Knowledge Guidance

After detecting a target object, a service robot must approach the target object to perform the associated service task. In active object detection (AOD) tasks, effective feature information representation and comprehensive action execution strategies are crucial. Currently, most AOD tasks are accomplished by traditional reinforcement learning algorithms, but there are still problems such as high task failure rates and model training efficiency. To solve these problems, this paper proposes a combined data-driven and knowledge-guided solution. First, semantic information features, depth information features and target object bounding box information are used as inputs to comprehensively represent feature information. Second, a policy network is constructed based on the proximal policy optimizaton (PPO) algorithm. The reward value is set according to the robot′s action, the position of the bounding box, and the distance to the target object, and then applied to the robot′s training process. Finally, the knowledge of the path experience in the task, the robot′s collision avoidance ability and the prediction of target object loss are combined to guide the robot′s behavior, and a comprehensive decision model is proposed to enable the robot to make the best decision. Relevant experiments were conducted on an active vision dataset. The robot achieves an average success rate of 91.36% and an average step size of 9.3631 in performing the AOD task in the test scenes, which verifies the effectiveness of the proposed scheme.