ON DATA-BASED MAPPING AND NAVIGATION OF UNMANNED GROUND VEHICLES

Abstract

Unmanned ground vehicles (UGVs) have seen tremendous advancement in their capabilities and applications in the past two decades. With several key algorithmic and hardware breakthroughs and advancements in deep learning, UGVs are quickly becoming ubiquitous (finding applications as self-driving cars, for remote site inspections, in hospitals and shopping malls, among several others). Motivated by their large-scale adoption, this dissertation aims to enable the navigation of UGVs in complex environments. In this dissertation, a supervised learning-based navigation algorithm that utilizes model predictive control (MPC) for providing training data is developed. Improving MPC performance by data-based modelling of complex vehicle dynamics is then addressed. Finally, this dissertation deals with detecting and registering transparent objects that may deteriorate navigation performance.

Navigation in dynamic environments poses unique challenges, particularly due to the limited knowledge of the decisions made by other agents and their objectives. In this dissertation, a solution that utilizes an MPC-based planner as an \textit{expert} to generate high-quality motion commands for a car-like robot operating in a simulated dynamic environment is proposed. These commands are then used to train a deep neural network, which learns to navigate. The deep learning-based planner is further enhanced with safety margins to improve its effectiveness in collision avoidance. The performance of the proposed method through simulations and real-world experiments, demonstrating its superiority in terms of obstacle avoidance and successful mission completion is showcased. This research has practical implications for the development of safer and more efficient autonomous vehicles.

Many real-world applications rely on MPC to control UGVs due to its safety guarantees and constraint satisfaction properties. However, the performance of such MPC-based solutions is heavily reliant on the accuracy of the motion model. This dissertation addresses this challenge by exploring a data-based approach to discovering vehicle dynamics. Unlike existing physics-based models that require extensive testing setups and manual tuning for new platforms and driving surfaces, our approach leverages the universal differential equations (UDEs) framework to identify unknown dynamics from vehicle data. This innovative approach, which does not make assumptions about the unknown dynamics terms and directly models the vector field, is then deployed to showcase its efficacy. This research opens up new possibilities for more accurate and adaptable motion models for UGVs.

With the increasing adoption of glass and other transparent materials, UGVS must be able to detect and register them for reliable navigation. Unfortunately, such objects are not easily detected by LiDARs and cameras. In this dissertation, algorithms for detecting and including glass objects in a Graph SLAM framework were studied. A simple and computationally inexpensive glass detection scheme to detect glass objects is utilized. The methodology to incorporate the identified objects into the occupancy grid maintained by such a framework is the presented. The issue of \textit{drift accumulation} that can affect mapping performance when operating in large environments is also addressed.