Behavior Cloning-based Active Scene Recognition via Generated Expert Data with Revision and Prediction for Domestic Robots

Given the limitations of current methods in terms of accuracy and efficiency for robot scene recognition (SR) in domestic environments, this paper proposes an active scene recognition approach (ASR) that allows the robot to recognize scenes correctly using less images, even when the robot's position and observation direction are uncertain. ASR includes a behavior cloning-based action classification model, which can adjust the robot view actively to capture beneficial images for scene recognition. To address the lack of essential expert data for training the action model, we introduce an expert data generation method that avoids time-consuming and inefficient manual data collection. Additionally, we present a multi-view scene recognition method to handle the multiple images resulting from view changes. This method includes a scene recognition model that scores each image and a revision and prediction method to mitigate the compounding error introduced by behavior cloning as well as output the finial recognition result. We conducted numerous comparative experiments and an ablation study in various domestic environments using a publicly simulated platform to validate our ASR method. The experimental results demonstrate that our proposed approach outperforms state-of-the-art methods in terms of both accuracy and efficiency for scene recognition. Furthermore, our method, trained in simulated environments, demonstrates excellent generalization capabilities, allowing it to be directly transferred to the real world without the need for fine-tuning. When deployed on a TurtleBot 4 robot, it achieves precise and efficient scene recognition in diverse real-world environments.