Resilient Large-scale Informative Path Planning for Autonomous Robotic Exploration
Loading...
Authors
Dang, Anh Tung
Issue Date
2020
Type
Dissertation
Language
Keywords
Aerial robot , Autonomous exploration , Autonomous robot , Legged robot , Subterranean robotics , Underground environments
Alternative Title
Abstract
Autonomous robotic exploration is expanding into an ever-increasing set of critical applications including surveillance, search and rescue, as well as commercial activities. Nevertheless, a variety of environments remain challenging for robotic entry and resilient autonomy. For instance, a search-and-rescue mission in a collapsed underground mine imposes major difficulties for robotic autonomy due to its large-scale geometrically-constrained structures, perceptually-degraded conditions, lack of communication infrastructure, and diverse terrain. Such settings push the limits of the current state-of-the-art in exploration planning despite its successes in smaller and rather structured environments. Simultaneously, autonomous exploration missions often require the gathering of other relevant information (e.g., visual images of objects of interest) in addition to the primary mapping task. Exploration planning algorithms, however, are usually agnostic to the specific visual importance of different objects and entities in the environment, which can lead to sparse and unfocused data collection. Furthermore, planning in such cases assumes the availability of precise localization and mapping for collision avoidance and information gain calculations, which may be very challenging to achieve in GPS-denied, large-scale and perceptually-degraded environments. In response to these facts, the focus of this study is on autonomous-informative exploration in large-scale and high-risk challenging environments.Part A of this dissertation aims to tackle the exploration problem in large-scale underground settings. We address it from a holistic perspective, providing a versatile and resilient solution working across different challenging conditions. A graph-based bifurcated local-global planning framework is proposed and embeds the ability to plan efficiently locally, while simultaneously facilitating global re-positioning to unexplored parts of the map from which the exploration can continue. We further develop a variant of this method focused on aerial robots for faster exploration exploiting motion primitives that respect the dynamics of MAVs. The proposed framework, involving both contributed planning algorithms, also abstracts the interface between the planners and the robot controller, which allows it to be integrated seamlessly on different robotic platforms. The new planners were field demonstrated successfully inside multiple subterranean environments across the U.S. and Switzerland using a diverse set of flying robots and legged systems. Notably, these planners have been successfully deployed in the DARPA Subterranean Challenge and have been the full planning solution employed by the CERBERUS Team.Part B focuses on the problem of informative exploration path planning for aerial robots, while also being cognizant of salient areas in the scene utilizing low-level visual cues that act as intrinsic motivations, namely saliency and anomaly detection. The planner uses color images to first detect salient areas and then allocates more and high quality observations towards those regions, presenting an effective and reliable approach for focused information sampling within an exploration mission. The proposed planning pipeline follows a nested two-layer optimization approach. In the first step, it considers exploration and mapping as the sole objective and generates a reference path for exploration. Given the reference path, the second layer re-plans another path that preserves the same exploration gain as the first, but re-optimizes the robot's viewpoints along the path to account for the intrinsic information gathering objective. Through simulation and experimental studies, the proposed planning method demonstrated significant improvement in the quality of visual data collected during the exploration mission which are more focused and more rich towards visually informative areas.Part C of the dissertation expands the research into a learning-based planning approach. We use an imitation learning method to compress the behavior of the local level of the graph-based expert planner developed in the first part of this thesis, into a single neural network model. The training happens only in simulation, while the method is deployed in real-life, and it provides a rather lightweight policy that - in the environments considered - has a performance analogous to that of the expert planner but with a fraction of computation cost. This approach opens a new frontier for planning under extreme conditions with less computation, especially in large-scale and perceptually-degraded environments where localization and mapping are often expensive and difficult to achieve. As a result, during field deployments inside a real underground mine, the planner was able to provide safe and efficient exploratory paths to progressively navigate the robot towards unknown space in the environment without utilizing a real-time map of the environment and in a computationally lightweight manner.In conclusion, this thesis approaches the problems of large-scale and long-term exploration in high-risk and complex settings such as - but not limited to - subterranean environments. It approaches this problem in three main ways, namely a) bifurcated methods for global-local exploration to ensure efficiency but also capacity to handle kilometers-long settings, b) multi-objective optimization to ensure autonomous mapping alongside focused search towards salient and anomalous areas, as well as c) a new learning paradigm to deliver a similar planning behavior but without the need to maintain a consistent map of the environment and at a fraction of the computational cost. This work also puts effort in the field verification, not only as a way to demonstrate the achieved performance, but as a process to acquire new valuable lessons for the research directions that need to be explored. In fact, our core contributions in subterranean exploration involve ideas understood and identified based on conclusions drawn from field experiments in a multitude of mines, urban underground infrastructure and cavern-like settings, where we saw the limitations of traditional state-of-the-art methods in exploration path planning.