Deep Learning-Powered Natural Language Interface in Intelligent Pest Extermination Robotics

Abstract

To address the low efficiency and inconvenient human-computer interaction of traditional pest control methods, this study designs and implements a natural language interaction system for intelligent pest control robots integrating a Vision-Language-Action (VLA) model. The hardware platform integrates an RGB-D camera and a precision spraying module. The software architecture employs FiS-VLA as the decision-making core, constructing a fast-slow dual-system reasoning mechanism. The perception layer uses a ViT and LLaMA-3 fusion encoder, while the execution layer achieves precise control from discrete commands to continuous actions through action tokenization and the flow matching model. Experiments show that, compared to baseline models like RT-2, our system achieves a significant increase in task success rate on a custom pest control command dataset. Notably, the integration of the RoboRefer module for 3D spatial reference parsing reduces the L1 error of nest localization to within 5cm, and the end-to-end system response latency reaches 200Hz under the fast system. This research confirms that the proposed VLA-NLP fusion framework significantly enhances the autonomous operational efficacy and intuitive human-robot interaction of pest control robots in complex, unstructured environments, providing an effective paradigm for developing next-generation agricultural intelligent equipment with open-world generalization capabilities.