Resilience in Control and Motion Planning for Autonomous robots

Abstract

Recent advances in autonomy research have enabled the widespread adoption of robots in multiple applications including for subterranean exploration, construction, agriculture, parcel delivery, and forestry. However, instilling reliability and resilience in autonomous robotic operations in a diverse set of challenging, geometrically complex, and perceptually-degraded environments remains demanding. Therefore, the goal of this thesis is to study core elements of the Science of Resilient Robotic Autonomy from several perspectives to pave the way for the resilient design paradigm in robotics. In pursuit of this goal, this thesis encompasses three distinct parts that address resilience in control, motion planning, and robotic systems. Part I contains research on the design, modeling, and control of a new type of aerial robot and a survey on the application and design considerations of Model Predictive Control (MPC) for aerial robots. Part II presents research on resilient learning-based navigation methods for autonomous robots. Part III discusses resilience in autonomous robotic systems, including an individual aerial robot for cave exploration and eventually a robotic system-of-systems in the DARPA Subterranean Challenge Final Event through Team CERBERUS. This thesis presents seven novel contributions allowing us to break new ground in the Science of Resilient Robotic Autonomy. Three of these are based on articles submitted or published in peer-review journals, and four are based on articles published in peer-reviewed conference proceedings. Other publications that the candidate contributed during this PhD (but are not directly related to the content of the chapters) are also listed. The first chapter discusses the system design, modeling, and control of a novel aerial robotic system. This novel robot design offers the potential to simultaneously carry a significant payload (including sensing and processing units), perform forceful physical interaction, and morph its shape in a versatile manner in order to negotiate narrow areas. A hybrid modeling framework including Free-flight and Aerial Manipulation modes is proposed to model the system and respective controllers are designed for both operating modes with stability guarantees provided by Lyapunov theory and numerically verified with reachability analysis. We demonstrate the stability and performance of a prototype system in a series of experimental studies including a task of valve rotation, a pick-and-release task, and the verification of load oscillation suppression. The next chapter encompasses a comprehensive survey of the utilization and design of MPC for Micro Aerial Vehicles (MAVs). Our study delves into methods considering the free-flight dynamics of robots, both linear and nonlinear, while accounting for state and input constraints. We also survey MPC approaches in physical interaction and load transportation tasks, fault-tolerant control schemes, and research combining MPC with reinforcement learning techniques. Additionally, we present selective simulation results that offer design guidelines for choosing between linear and nonlinear schemes, tuning the prediction horizon, emphasizing the significance of disturbance observer-based offset-free tracking, and highlighting the inherent robustness of such methods to parameter uncertainty. Finally, a set of open-source code packages that provide readily available functionality for MPC deployment onboard MAVs are categorized. Part II focuses on designing resilient learning-based navigation methods that allow the robots to operate autonomously without relying on a global map of the environment or the robot’s position information, subjected to noisy onboard proprioceptive sensor input and uncertain robot’s partial state estimate (excluding the robot’s position). Specifically, the work in Chapter 4 contributes a navigation method that allows safe uncertainty-aware 2D navigation. At the core of the method is a deep neural network taking as inputs a) the current depth image, b) the robot’s partial state, and c) a motion primitives library in velocity-steering angle space to predict the collision score of each action sequence in the motion primitives library. Notably, we utilized the Unscented Transform over the robot’s partial state and the Monte Carlo dropout method to derive uncertainty-aware collision costs that can be estimated efficiently utilizing the parallel computing capability of modern Graphical Process Unit (GPU). These collision costs can then be used in addition to a direction command given by a high-level planner to choose the best action to be executed in a receding horizon manner. A set of simulation and experimental studies, including field deployment in an underground mine, is conducted to evaluate the quality of the prediction network and the performance of the proposed method. Chapter 5 presents major extensions to Chapter 4: First, it is about introducing visual attention-aware navigation into the framework through the Information gain Prediction Network. Second, we extend the previous work from 2D to 3D navigation and further utilize a deep ensembles method for the neural network’s epistemic uncertainty estimation. Third, a new set of simulations and real-world experiments are conducted to verify the proposed uncertainty-aware and visuallyattentive framework. In particular, more thorough simulation studies are conducted to demonstrate the performance of our method against noisy inputs including the robot’s velocity estimate and the depth image. Moreover, simulation results with different sources of visual attention are performed to illustrate the advantages of our visually-attentive navigation method compared to other baselines. Finally, realworld experiments in a diverse set of environments including forests, an industrial facility, and cluttered corridors in office buildings (with a reference speed up to 2.5 m/s) are presented. While the end-to-end approaches in Chapter 4 and Chapter 5 can transfer well to the real system with the help of a fairly expensive image pre-processing step to close the sim-to-real gap, they don’t allow the assimilation of real-world exteroceptive sensor data into the training pipeline. Moreover, encoding of hard-to-perceive thin obstacles in the latent vector for collision prediction tasks cannot be enforced explicitly with the works in Chapter 4 and Chapter 5. Chapter 6 addresses these limitations by proposing a modularized learning-based method for aerial robots navigating cluttered environments containing hard-to-perceive thin obstacles without assuming access to a map or the full pose estimation of the robot. It is built upon a semantically-enhanced Variational Autoencoder that is trained with both real-world and simulated depth images to compress the input data, while preserving semantically-labeled thin obstacles and handling invalid pixels in the depth sensor’s output. This compressed representation and the robot’s partial state are then utilized to train an uncertainty-aware 3D Collision Prediction Network with simulated collision data, as in Chapter 5, to predict collision scores for candidate action sequences in a motion primitives library. We conducted a set of simulation and experimental studies in cluttered environments with various sizes and types of obstacles, including multiple hard-to-perceive thin objects, to evaluate the performance of the proposed method and demonstrate its benefits against an end-to-end trained baseline. Chapter 7 contributes a complete design of an aerial robot capable of autonomous exploration of cave networks. The system relies on two core functionalities for multimodal localization and mapping, as well as path planning for enabling resilient autonomy in such diverse and challenging environments. Specifically, the system utilizes a multi-modal sensor suite including LiDAR, visible-light and thermal cameras, and inertial sensing for robust and resourceful perception in the GPS-denied, dark, often obscurants-filled and geometrically complex cave settings. Building on top of its perception capabilities, the system implements a graph-based exploration path planning algorithm tailored to the topological patterns observed in caves. To evaluate the proposed solution, a set of simulation studies and field experiments were conducted and presented in detail. Chapter 8 details the employed robotic system-of-systems of Team CERBERUS participating in the DARPA Subterranean (SubT) Challenge and specifically towards the Final Event in 2021 in which CERBERUS won this prestigious international competition. The team of robots includes legged and flying systems, including collision-tolerant designs, further enhanced with a roving platform. The implemented autonomy components, including multi-modal perception and pathplanning, navigation and control of legged robots, and the automated artifact detection and scoring systems, are detailed alongside the operator interfaces for exerting high-level control.We then presented the Team’s performance in the winning Prize Round of the challenge, followed by the critical lessons learned from these experiences to design a resilient autonomous robotic system-of-systems. My contributions towards CERBERUS’s system-of-systems (including methodology, system integration, and support in field deployments) are also discussed.