Learn to Trace Odors: Robotic Odor Source Localization via Deep Learning Methods with Real-world Experiments

Abstract

This paper presents an olfactory-based navigation algorithm via deep learning (DL) methods. The object is to obtain a neural network that navigates a mobile robot to find an odor source without explicating search strategies. Two deep neural networks (DNNs), including feedforward and long short-term memory neural networks (FNN and LSTM), are devised to calculate robot heading commands based on onboard sensor readings. The training dataset is obtained by implementing traditional olfactory-based navigation algorithms, namely moth-inspired and Bayesian-inference methods, on a mobile robot in hundreds of odor source localization (OSL) tests. After the supervised learning, DNNs are validated in real-world experiments with unseen odor source locations and airflow fields. Experiment results show that both FNN and LSTM can imitate the moth-inspired method but cannot effectively learn the complex Bayesian-inference method. In terms of the averaged search time in repeated tests, the proposed FNN and LSTM outperform the Bayesian-inference method by 18% and 14%, respectively, and both networks achieve a comparable search performance with the moth-inspired method.