Agile autonomy: learning tightly-coupled perception-action for high-speed quadrotor flight in the wild
Antonio Loquercio · Zurich Open Repository and Archive (University of Zurich) · 2021
The robot revolution has arrived", says the stunning title of a 2020 article in the National Geographic.Over the course of the last decades, autonomous robots have had a tremendous impact on our global economy.Indeed, machines now perform all sorts of tasks: they take inventory and clean floors in big stores; they shelve goods and fetch them for mailing in warehouses; they patrol borders; and they help children with autism.In the majority of those applications, they co-exist and interact with their surroundings, being that humans or other autonomous systems, to create an ecosystem that maximizes efficiency and productivity.Yet, a major challenge for autonomous systems to operate in unconstrained settings is to cope with the constant variability and uncertainty of the real world.This challenge currently limits the application of robots to structured environments, where they can be closely monitored by expensive suites of sensors and/or human operators.One way to tackle this challenge is exploiting the synergy between perception and action which characterizes natural and artificial agents alike.For embodied agents, action controls the amount of information coming from sensory data, and perception guides and provides feedback to action.While this seamless integration of sensing and control a fundamental feature of biological agents [91], I argue that a tight perception-action loop is fundamental in enriching the autonomy of artificial agents.My thesis investigates this question in the context of high-speed agile quadrotor flight.Given their agility, limited cost, and widespread availability, quadrotors are the perfect platform to demonstrate the advantages of a tightly coupled perception and action loop.To date, only human pilots can fully exploit their capabilities in unconstrained settings.Autonomous operation has been limited to constrained environments and/or low speeds.Having reached human-level performance on several tasks, data-driven algorithms, e.g.neural networks, represent the ideal candidate to enhance drones' autonomy.However, this computational tool has mainly proven its impact on disembodied datasets or in very controlled conditions, e.g. in standardized visual recognition tasks or games.Therefore, exploiting their potential for high-speed aerial robotics in a constantly changing and uncertain world poses both technical and fundamental challenges.This thesis presents algorithms for tightly-coupled robotic perception and action in unstructured environments, where a robot can only rely on on-board sensing and computation to accomplish a mission.In the following, we define a policy as tightly-coupled if it directly goes from onboard sensor observations to actions without intermediate steps.Doing so creates a synergy between perception and action, and it enables end-to-end optimization to the downstream task.To design such policy, I explore the possibilities of data-driven algorithms.Among others, this thesis provides contributions to achieving high-speed agile quadrotor flight in the wild, pushing iii Abstract the platform closer to its physical limits than what was possible before.Specifically, I study the problem of navigation in natural and man-made (hence in the wild) previously unknown environments with only on-board sensing and computations (neither a GPS nor external links are available to the robot).To achieve this goal, I introduce the idea of sensing and action abstraction to achieve knowledge transfer between different platforms, e.g. from ground to aerial robots, and domain, e.g. from simulation to the real world.This thesis also presents novel ways to estimate uncertainty in neural network predictions, which open the door to integrating data-driven methods in robotics in a safe and accountable fashion.Finally, this thesis also explores the possibilities that the embodied nature of robots brings to tackle complex perception problems, both for low-level tasks such as depth estimation, or high-level tasks (e.g.moving object detection).Overall, the contributions of this work aim to answer the question Why learning tightly-coupled sensorimotor policies in robotics?In the following is a list of contributions:• The introduction of data-driven algorithms to the problem of high-speed agile quadrotor flight in unstructured environments with only on-board sensing and computation.