LibSmart: an Open-Source Tool for Simple Integration of Deep Learning into Intelligent Robotic Systems

Abstract

Intelligent robotic systems can be empowered by advanced deep learning techniques. Robotic operations such as object recognition are well investigated by researchers involved in machine learning. However, these solutions have often led to ad-hoc implementation in experimental settings. Less reported is systematic implementation of deep learning models in industrial robots. The lack of standard implementation platforms has impeded widespread use of deep learning modules in industrial robots. It is of great importance to have development platforms that can coordinate several deep learning modules of a complex system. In this paper, a scalable deep-learning friendly robot task organization system named libSmart is introduced. Similar to ROS, the architecture of the proposed system allows users to plug and play various devices but the proposed architecture is also highly compatible with deep learning modules. Specifically, the deployment of deep learning models is handled using a novel data graph method with distributed computing. In this way, the computationally expensive training and inferencing processes of deep learning models can be handled with isolated accelerating hardware to reduce the overall system latency. Successful implementation of simultaneous object recognition and pose estimation by an industrial robot has been presented as a case study. The proposed system is open source for all users to build their own intelligent systems with customized deep-learning models. (https://github.com/RustIron/libSmart.git).