A Review of Recent Research On Deep Learning in Robotics
The science of training large artificial neural networks is known as deep learning. DNNs can have hundreds of millions of parameters, allowing them to model complex functions like nonlinear dynamics. They create compact state representations from raw, high-dimensional, multimodal sensor data found in robotic systems, and unlike many machine learning methods, they don't require a human expert to hand-engineer feature vectors from sensor data at design time. Deep learning advances have sparked a flurry of research in the application of deep artificial neural networks to robotic systems over the last decade, with at least 30 papers published on the topic between 2014 and now. Using current research as examples, this review discusses the applications, benefits, and limitations of deep learning in relation to physical robotic systems. Its goal is to inform the broader robotics community about recent advances and to pique interest in and application of deep learning in robotics.