A Lightweight Operation Mode Decision Method for Cleaning Robots Driven by Garbage Attributes Perception

Abstract

The intelligent operation mode decision scheme has been proved to be a promising solution for enhancing the cleaning performance of cleaning robots. In this paper, we propose a lightweight operation mode decision method for cleaning robots, driven by garbage attributes perception. The method aims to enable cleaning robots to intelligently select the most appropriate operation mode when dealing with different types of garbage, thereby improving their cleaning efficiency. Specifically, we propose a lightweight garbage attributes extraction network (LGAE-Net) based on inverted residuals, which adopts the proposed deep dynamic attention convolution (DDA-Conv) as its basic structure and adaptively extracts share features of attributes while reducing computational complexity. Then, the network adopts a multi-label architecture to predict multiple attributes, and a dynamic weighting joint learning strategy is introduced to alleviate the problem of imbalanced difficulty in attribute learning. Finally, based on the extraction of attributes, a decision module for operation mode is constructed. This module enables fast decision from the perception of attributes to the selection of an operation mode. Our proposed method achieves 94.26% decision accuracy on the test dataset, and the single-sheet recognition rate is only 1.63 msec. In addition, it maintains a parameter count (Params(M)) of approximately 2.97M and floating-point operations (FLOPs(M)) of only 94.91M, which reflects the excellent balance of accuracy and efficiency of the method, and can meet the real-time requirements of cleaning robots.

Keywords

• Attribute learning
• cleaning robots
• lightweight network
• multiple modes decision