Eusocial insect colonies have often been imagined as “superorganisms” exhibiting tight homeostasis at the colony level. However, colonies lack the tight spatial and organizational integration that many multicellular, unitary organisms exhibit. Precise regulation requires rapid feedback, which is often not possible when nestmates are distributed across space, making decisions asynchronously. Thus, one should expect poorer regulation in superorganisms than unitary organisms.Here, I investigate aspects of regulation in collective foraging behaviors that involve both slow and rapid feedback processes. In Chapter 2, I examine a tightly coupled system with near-instantaneous signaling: teams of weaver ants cooperating to transport massive prey items back to their nest. I discover that over an extreme range of scenarios—even up vertical surfaces—the efficiency per transporter remains constant. My results suggest that weaver ant colonies are maximizing their total intake rate by regulating the allocation of transporters among loads. This is an exception that “proves the rule;” the ant teams are recapitulating the physical integration of unitary organisms. Next, I focus on a process with greater informational constraints, with loose temporal and spatial integration. In Chapter 3, I measure the ability of solitarily foraging Ectatomma ruidum colonies to balance their collection of protein and carbohydrates given different nutritional environments. Previous research has found that ant species can precisely collect a near-constant ratio between these two macronutrients, but I discover these studies were using flawed statistical approaches. By developing a quantitative measure of regulatory effect size, I show that colonies of E. ruidum are relatively insensitive to small differences in food source nutritional content, contrary to previously published claims. In Chapter 4, I design an automated, micro-RFID ant tracking system to investigate how the foraging behavior of individuals integrates into colony-level nutrient collection. I discover that spatial fidelity to food resources, not individual specialization on particular nutrient types, best predicts individual forager behavior. These findings contradict previously published experiments that did not use rigorous quantitative measures of specialization and confounded the effects of task type and resource location.

Multi-segment manipulators and mobile robot collectives are examples of multi-agent robotic systems, in which each segment or robot can be considered an agent. Fundamental motion control problems for such systems include the stabilization of one or more agents to target configurations or trajectories while preventing inter-agent collisions, agent collisions with obstacles, and deadlocks. Despite extensive research on these control problems, there are still challenges in designing controllers that (1) are scalable with the number of agents; (2) have theoretical guarantees on collision-free agent navigation; and (3) can be used when the states of the agents and the environment are only partially observable. Existing centralized and distributed control architectures have limited scalability due to their computational complexity and communication requirements, while decentralized control architectures are often effective only under impractical assumptions that do not hold in real-world implementations. The main objective of this dissertation is to develop and evaluate decentralized approaches for multi-agent motion control that enable agents to use their onboard sensors and computational resources to decide how to move through their environment, with limited or absent inter-agent communication and external supervision. Specifically, control approaches are designed for multi-segment manipulators and mobile robot collectives to achieve position and pose (position and orientation) stabilization, trajectory tracking, and collision and deadlock avoidance. These control approaches are validated in both simulations and physical experiments to show that they can be implemented in real-time while remaining computationally tractable. First, kinematic controllers are proposed for position stabilization and trajectory tracking control of two- or three-dimensional hyper-redundant multi-segment manipulators. Next, robust and gradient-based feedback controllers are presented for individual holonomic and nonholonomic mobile robots that achieve position stabilization, trajectory tracking control, and obstacle avoidance. Then, nonlinear Model Predictive Control methods are developed for collision-free, deadlock-free pose stabilization and trajectory tracking control of multiple nonholonomic mobile robots in known and unknown environments with obstacles, both static and dynamic. Finally, a feedforward proportional-derivative controller is defined for collision-free velocity tracking of a moving ground target by multiple unmanned aerial vehicles.

This work has improved the quality of the solution to the sparse rewards problemby combining reinforcement learning (RL) with knowledge-rich planning. Classical methods for coping with sparse rewards during reinforcement learning modify the reward landscape so as to better guide the learner. In contrast, this work combines RL with a planner in order to utilize other information about the environment. As the scope for representing environmental information is limited in RL, this work has conflated a model-free learning algorithm – temporal difference (TD) learning – with a Hierarchical Task Network (HTN) planner to accommodate rich environmental information in the algorithm. In the perpetual sparse rewards problem, rewards reemerge after being collected within a fixed interval of time, culminating in a lack of a well-defined goal state as an exit condition to the problem. Incorporating planning in the learning algorithm not only improves the quality of the solution, but the algorithm also avoids the ambiguity of incorporating a goal of maximizing profit while using only a planning algorithm to solve this problem. Upon occasionally using the HTN planner, this algorithm provides the necessary tweak toward the optimal solution. In this work, I have demonstrated an on-policy algorithm that has improved the quality of the solution over vanilla reinforcement learning. The objective of this work has been to observe the capacity of the synthesized algorithm in finding optimal policies to maximize rewards, awareness of the environment, and the awareness of the presence of other agents in the vicinity.

For a system of autonomous vehicles functioning together in a traffic scene, 3Dunderstanding of participants in the field of view or surrounding is very essential for assessing the safety operation of the involved. This problem can be decomposed into online pose and shape estimation, which has been a core research area of computer vision for over a decade now. This work is an add-on to support and improve the joint estimate of the pose and shape of vehicles from monocular cameras. The objective of jointly estimating the vehicle pose and shape online is enabled by what is called an offline reconstruction pipeline. In the offline reconstruction step, an approach to obtain the vehicle 3D shape with keypoints labeled is formulated. This work proposes a multi-view reconstruction pipeline using images and masks which can create an approximate shape of vehicles and can be used as a shape prior. Then a 3D model-fitting optimization approach to refine the shape prior using high quality computer-aided design (CAD) models of vehicles is developed. A dataset of such 3D vehicles with 20 keypoints annotated is prepared and call it the AvaCAR dataset. The AvaCAR dataset can be used to estimate the vehicle shape and pose, without having the need to collect significant amounts of data needed for adequate training of a neural network. The online reconstruction can use this synthesis dataset to generate novel viewpoints and simultaneously train a neural network for pose and shape estimation. Most methods in the current literature using deep neural networks, that are trained to estimate pose of the object from a single image, are inherently biased to the viewpoint of the images used. This approach aims at addressing these existing limitations in the current method by delivering the online estimation a shape prior which can generate novel views to account for the bias due to viewpoint. The dataset is provided with ground truth extrinsic parameters and the compact vector based shape representations which along with the multi-view dataset can be used to efficiently trained neural networks for vehicle pose and shape estimation. The vehicles in this library are evaluated with some standard metrics to assure they are capable of aiding online estimation and model based tracking.

A Graph Neural Network (GNN) is a type of neural network architecture that operates on data consisting of objects and their relationships, which are represented by a graph. Within the graph, nodes represent objects and edges represent associations between those objects. The representation of relationships and correlations between data is unique to graph structures. GNNs exploit this feature of graphs by augmenting both forms of data, individual and relational, and have been designed to allow for communication and sharing of data within each neural network layer. These benefits allow each node to have an enriched perspective, or a better understanding, of its neighbouring nodes and its connections to those nodes. The ability of GNNs to efficiently process high-dimensional node data and multi-faceted relationships among nodes gives them advantages over neural network architectures such as Convolutional Neural Networks (CNNs) that do not implicitly handle relational data. These quintessential characteristics of GNN models make them suitable for solving problems in which the correspondences among input data are needed to produce an accurate and precise representation of these data. GNN frameworks may significantly improve existing communication and control techniques for multi-agent tasks by implicitly representing not only information associated with the individual agents, such as agent position, velocity, and camera data, but also their relationships with one another, such as distances between the agents and their ability to communicate with one another. One such task is a multi-agent navigation problem in which the agents must coordinate with one another in a decentralized manner, using proximity sensors only, to navigate safely to their intended goal positions in the environment without collisions or deadlocks. The contribution of this thesis is the design of an end-to-end decentralized control scheme for multi-agent navigation that utilizes GNNs to prevent inter-agent collisions and deadlocks. The contributions consist of the development, simulation and evaluation of the performance of an advantage actor-critic (A2C) reinforcement learning algorithm that employs actor and critic networks for training that simultaneously approximate the policy function and value function, respectively. These networks are implemented using GNN frameworks for navigation by groups of 3, 5, 10 and 15 agents in simulated two-dimensional environments. It is observed that in $40\%$ to $50\%$ of the simulation trials, between 70$\%$ to 80$\%$ of the agents reach their goal positions without colliding with other agents or becoming trapped in deadlocks. The model is also compared to a random run simulation, where actions are chosen randomly for the agents and observe that the model performs notably well for smaller groups of agents.

Chemical Reaction Networks (CRNs) provide a useful framework for modeling andcontrolling large numbers of agents that undergo stochastic transitions between a set of states in a manner similar to chemical compounds. By utilizing CRN models to design agent control policies, some of the computational challenges in the coordination of multi-agent systems can be overcome. In this thesis, a CRN model is developed that defines agent control policies for a multi-agent construction task. The use of surface CRNs to overcome the tradeoff between speed and accuracy of task performance is explained. The computational difficulties involved in coordinating multiple agents to complete collective construction tasks is then discussed. A method for stochastic task and motion planning (TAMP) is proposed to explain how a TAMP solver can be applied with CRNs to coordinate multiple agents. This work defines a collective construction scenario in which a group of noncommunicating agents must rearrange blocks on a discrete domain with obstacles into a predefined target distribution. Four different construction tasks are considered with 10, 20, 30, or 40 blocks, and a simulation of each scenario with 2, 4, 6, or 8 agents is performed. As the number of blocks increases, the construction problem becomes more complex, and a given population of agents requires more time to complete the task. Populations of fewer than 8 agents are unable to solve the 30-block and 40-block problems in the allotted simulation time, suggesting an inflection point for computational feasibility, implying that beyond that point the solution times for fewer than 8 agents would be expected to increase significantly. For a group of 8 agents, the time to complete the task generally increases as the number of blocks increases, except for the 30-block problem, which has specifications that make the task slightly easier for the agents to complete compared to the 20-block problem. For the 10-block and 20- block problems, the time to complete the task decreases as the number of agents increases; however, the marginal effect of each additional two agents on this time decreases. This can be explained through the pigeonhole principle: since there area finite number of states, when the number of agents is greater than the number of available spaces, deadlocks start to occur and the expectation is that the overall solution time to tend to infinity.

Building and optimizing a design for deformable media can be extremely costly. However, granular scaling laws enable the ability to predict system velocity and mobility power consumption by testing at a smaller scale in the same environment. The validity of the granular scaling laws for arbitrarily shaped wheels and screws were evaluated in materials like silica sand and BP-1, a lunar simulant. Different wheel geometries, such as non-grousered and straight and bihelically grousered wheels were created and tested using 3D printed technologies. Using the granular scaling laws and the empirical data from initial experiments, power and velocity were predicted for a larger scaled version then experimentally validated on a dynamic mobility platform. Working with granular media has high variability in material properties depending on initial environmental conditions, so particular emphasis was placed on consistency in the testing methodology. Through experiments, these scaling laws have been validated with defined use cases and limitations.

Multibody Dynamic (MBD) models are important tools in motion analysis and are used to represent and accurately predict the behavior of systems in the real-world. These models have a range of applications, including the stowage and deployment of flexible deployables on spacecraft, the dynamic response of vehicles in automotive design and crash testing, and mapping interactions of the human body. An accurate model can aid in the design of a system to ensure the system is effective and meets specified performance criteria when built. A model may have many design parameters, such as geometrical constraints and component mechanical properties, or controller parameters if the system uses an external controller. Varying these parameters and rerunning analyses by hand to find an ideal design can be time consuming for models that take hours or days to run. To reduce the amount of time required to find a set of parameters that produces a desired performance, optimization is necessary. Many papers have discussed methods for optimizing rigid and flexible MBD models, and separately their controllers, using both gradient-based and gradient-free algorithms. However, these optimization methods have not been used to optimize full-scale MBD models and their controllers simultaneously. This thesis presents a method for co-optimizing an MBD model and controller that allows for the flexibility to find model and controller-based solutions for systems with tightly coupled parameters. Specifically, the optimization is performed on a quadrotor drone MBD model undergoing disturbance from a slung load and its position controller to meet specified position error performance criteria. A gradient-free optimization algorithm and multiple objective approach is used due to the many local optima from the tradeoffs between the model and controller parameters. The thesis uses nine different quadrotor cases with three different position error formulations. The results are used to determine the effectiveness of the optimization and the ability to converge on a single optimal design. After reviewing the results, the optimization limitations are discussed as well as the ability to transition the optimization to work with different MBD models and their controllers.

Soft robots provide an additional measure of safety and compliance over traditionalrigid robots. Generally, control and modelling experiments take place using a motion capture system for measuring robot configuration. While accurate, motion capture systems are expensive and require re-calibration whenever the cameras are adjusted. While advances in soft sensors contribute to a potential solution to sensing outside of a lab environment, most of these sensing methods require the sensors to be embedded into the soft robot arm. In this work, a more practical sensing method is proposed using off-the-shelf sensors and a Robust Extended Kalman Filter based sensor fusion method. Inertial measurement unit sensors and wire draw sensors are used to accurately estimate the state of the robot. An explanation for the need for sensor fusion is included in this work. The sensor fusion state estimate is compared to a motion capture measurement along with the raw inertial measurement unit reading to verify the accuracy of the results. The potential for this sensing system is further validated through Linear Quadratic Gaussian control of the soft robot. The Robust Extended Kalman Filter based sensor fusion shows an error of less than one degree when compared to the motion capture system.

Despite the rapid adoption of robotics and machine learning in industry, their application to scientific studies remains under-explored. Combining industry-driven advances with scientific exploration provides new perspectives and a greater understanding of the planet and its environmental processes. Focusing on rock detection, mapping, and dynamics analysis, I present technical approaches and scientific results of developing robotics and machine learning technologies for geomorphology and seismic hazard analysis. I demonstrate an interdisciplinary research direction to push the frontiers of both robotics and geosciences, with potential translational contributions to commercial applications for hazard monitoring and prospecting. To understand the effects of rocky fault scarp development on rock trait distributions, I present a data-processing pipeline that utilizes unpiloted aerial vehicles (UAVs) and deep learning to segment densely distributed rocks in several orders of magnitude. Quantification and correlation analysis of rock trait distributions demonstrate a statistical approach for geomorphology studies. Fragile geological features such as precariously balanced rocks (PBRs) provide upper-bound ground motion constraints for hazard analysis. I develop an offboard method and onboard method as complementary to each other for PBR searching and mapping. Using deep learning, the offboard method segments PBRs in point clouds reconstructed from UAV surveys. The onboard method equips a UAV with edge-computing devices and stereo cameras, enabling onboard machine learning for real-time PBR search, detection, and mapping during surveillance. The offboard method provides an efficient solution to find PBR candidates in existing point clouds, which is useful for field reconnaissance. The onboard method emphasizes mapping individual PBRs for their complete visible surface features, such as basal contacts with pedestals–critical geometry to analyze fragility. After PBRs are mapped, I investigate PBR dynamics by building a virtual shake robot (VSR) that simulates ground motions to test PBR overturning. The VSR demonstrates that ground motion directions and niches are important factors determining PBR fragility, which were rarely considered in previous studies. The VSR also enables PBR large-displacement studies by tracking a toppled-PBR trajectory, presenting novel methods of rockfall hazard zoning. I build a real mini shake robot providing a reverse method to validate simulation experiments in the VSR.