Efficient Deployment of Deep Learning Models on Autonomous Robots in the ROS Environment

Abstract

Abstract Autonomous robots are often deployed in applications to continually monitor changing environments such as supermarket floors or inventory monitoring, patient monitoring, and autonomous driving. With the increasing use of deep learning techniques in robotics, a large number of robot manufacturing companies have started adopting deep learning techniques to improve the monitoring performance of autonomous robots. The Robot Operating System (ROS) is a widely used middleware platform for building autonomous robot applications. However, the deployment of deep learning models to autonomous robots using ROS remains an unexplored area of research. Most recent research has focused on using deep learning techniques to solve specific problems (e.g., shopping assistant robots, autopilot systems, automatic annotation of 3D maps for safe flight). However, integrating the data collection hardware (e.g., sensors) and deep learning models within ROS is difficult and expensive in terms of computational power, time, and energy (battery). To address these challenges, we have developed EasyDLROS, a novel framework for robust deployment of pre-trained deep learning models on robots. Our framework is open-source, independent of the underlying deep learning framework, and easy to deploy. To test the performance of EasyDLROS, we deployed seven pre-trained deep learning models for hazard detection on supermarkets floors in a simulated environment and evaluated their performances. Experimental results show that our framework successfully deploys the deep learning models on ROS environment.