Machine Learning for Autonomous Vehicle Navigation in Unstructured Environments

Abstract

As we advance towards creating vehicles and robots that can operate over a wide array of environments both indoors and outdoors, the emphasis in autonomous vehicle navigation is moving away from conventional control based architectures towards using machine learning models. These models are designed with the capability to learn from first principles and hence can be applied across a broad array of scenarios. In a review by Dong et al., the authors have noted that recent advancements in machine learning have improved the capabilities of autonomous navigational systems in detecting objects, understanding the semantic map of the environment, and planning and executing optimized path trajectories. However, the mechanisms for safety and human-like interpretability of these models requires further attention. Furthermore to improve and encourage further advancements in machine learning based autonomous navigation, the authors have proposed a standardization of datasets and benchmark environments. Deep learning models can also be used for robot navigation in agricultural settings wherein we can design conservation tillage equipment on top of autonomous field robots to detect row crops and end-rows using image-based classification and detection algorithms. For these deep learning based systems to work in outdoor agricultural settings, they must overcome environmental challenges such as varying terrain, weather, and illumination conditions.Researchers have also proposed a navigation model that is fully based on learning-based strategies compared to the traditional offline training method to improve robot navigation by reducing redundancy.