Robots are often used in long-duration scenarios, such as on the surface of Mars,where they may need to adapt to environmental changes. Typically, robots have been built specifically for single tasks, such as moving boxes in a warehouse or surveying construction sites. However, there is a modern trend away from human hand-engineering and toward robot learning. To this end, the ideal robot is not engineered,but automatically designed for a specific task. This thesis focuses on robots which learn path-planning algorithms for specific environments. Learning is accomplished via genetic programming. Path-planners are represented as Python code, which is optimized via Pareto evolution. These planners are encouraged to explore curiously and efficiently. This research asks the questions: “How can robots exhibit life-long learning where they adapt to changing environments in a robust way?”, and “How can robots learn to be curious?”.

A complex social system, whether artificial or natural, can possess its macroscopic properties as a collective, which may change in real time as a result of local behavioral interactions among a number of agents in it. If a reliable indicator is available to abstract the macrolevel states, decision makers could use it to take a proactive action, whenever needed, in order for the entire system to avoid unacceptable states or con-verge to desired ones. In realistic scenarios, however, there can be many challenges in learning a model of dynamic global states from interactions of agents, such as 1) high complexity of the system itself, 2) absence of holistic perception, 3) variability of group size, 4) biased observations on state space, and 5) identification of salient behavioral cues. In this dissertation, I introduce useful applications of macrostate estimation in complex multi-agent systems and explore effective deep learning frameworks to ad-dress the inherited challenges. First of all, Remote Teammate Localization (ReTLo)is developed in multi-robot teams, in which an individual robot can use its local interactions with a nearby robot as an information channel to estimate the holistic view of the group. Within the problem, I will show (a) learning a model of a modular team can generalize to all others to gain the global awareness of the team of variable sizes, and (b) active interactions are necessary to diversify training data and speed up the overall learning process. The complexity of the next focal system escalates to a colony of over 50 individual ants undergoing 18-day social stabilization since a chaotic event. I will utilize this natural platform to demonstrate, in contrast to (b), (c)monotonic samples only from “before chaos” can be sufficient to model the panicked society, and (d) the model can also be used to discover salient behaviors to precisely predict macrostates.

A swarm describes a group of interacting agents exhibiting complex collective behaviors. Higher-level behavioral patterns of the group are believed to emerge from simple low-level rules of decision making at the agent-level. With the potential application of swarms of aerial drones, underwater robots, and other multi-robot systems, there has been increasing interest in approaches for specifying complex, collective behavior for artificial swarms. Traditional methods for creating artificial multi-agent behaviors inspired by known swarms analyze the underlying dynamics and hand craft low-level control logics that constitute the emerging behaviors. Deep learning methods offered an approach to approximate the behaviors through optimization without much human intervention.

This thesis proposes a graph based neural network architecture, SwarmNet, for learning the swarming behaviors of multi-agent systems. Given observation of only the trajectories of an expert multi-agent system, the SwarmNet is able to learn sensible representations of the internal low-level interactions on top of being able to approximate the high-level behaviors and make long-term prediction of the motion of the system. Challenges in scaling the SwarmNet and graph neural networks in general are discussed in detail, along with measures to alleviate the scaling issue in generalization is proposed. Using the trained network as a control policy, it is shown that the combination of imitation learning and reinforcement learning improves the policy more efficiently. To some extent, it is shown that the low-level interactions are successfully identified and separated and that the separated functionality enables fine controlled custom training.

Accurate pose initialization and pose estimation are crucial requirements in on-orbit space assembly and various other autonomous on-orbit tasks. However, pose initialization and pose estimation are much more difficult to do accurately and consistently in space. This is primarily due to not only the variable lighting conditions present in space, but also the power requirements mandated by space-flyable hardware. This thesis investigates leveraging a deep learning approach for monocular one-shot pose initialization and pose estimation. A convolutional neural network was used to estimate the 6D pose of an assembly truss object. This network was trained by utilizing synthetic imagery generated from a simulation testbed. Furthermore, techniques to quantify model uncertainty of the deep learning model were investigated and applied in the task of in-space pose estimation and pose initialization. The feasibility of this approach on low-power computational platforms was also tested. The results demonstrate that accurate pose initialization and pose estimation can be conducted using a convolutional neural network. In addition, the results show that the model uncertainty can be obtained from the network. Lastly, the use of deep learning for pose initialization and pose estimation in addition with uncertainty quantification was demonstrated to be feasible on low-power compute platforms.

There are more than 20 active missions exploring planets and small bodies beyond Earth in our solar system today. Many more have completed their journeys or will soon begin. Each spacecraft has a suite of instruments and sensors that provide a treasure trove of data that scientists use to advance our understanding of the past, present, and future of the solar system and universe. As more missions come online and the volume of data increases, it becomes more difficult for scientists to analyze these complex data at the desired pace. There is a need for systems that can rapidly and intelligently extract information from planetary instrument datasets and prioritize the most promising, novel, or relevant observations for scientific analysis. Machine learning methods can serve this need in a variety of ways: by uncovering patterns or features of interest in large, complex datasets that are difficult for humans to analyze; by inspiring new hypotheses based on structure and patterns revealed in data; or by automating tedious or time-consuming tasks. In this dissertation, I present machine learning solutions to enhance the tactical planning process for the Mars Science Laboratory Curiosity rover and future tactically-planned missions, as well as the science analysis process for archived and ongoing orbital imaging investigations such as the High Resolution Imaging Science Experiment (HiRISE) at Mars. These include detecting novel geology in multispectral images and active nuclear spectroscopy data, analyzing the intrinsic variability in active nuclear spectroscopy data with respect to elemental geochemistry, automating tedious image review processes, and monitoring changes in surface features such as impact craters in orbital remote sensing images. Collectively, this dissertation shows how machine learning can be a powerful tool for facilitating scientific discovery during active exploration missions and in retrospective analysis of archived data.

This thesis presents a family of adaptive curvature methods for gradient-based stochastic optimization. In particular, a general algorithmic framework is introduced along with a practical implementation that yields an efficient, adaptive curvature gradient descent algorithm. To this end, a theoretical and practical link between curvature matrix estimation and shrinkage methods for covariance matrices is established. The use of shrinkage improves estimation accuracy of the curvature matrix when data samples are scarce. This thesis also introduce several insights that result in data- and computation-efficient update equations. Empirical results suggest that the proposed method compares favorably with existing second-order techniques based on the Fisher or Gauss-Newton and with adaptive stochastic gradient descent methods on both supervised and reinforcement learning tasks.

Automated driving systems are in an intensive research and development stage, and the companies developing these systems are targeting to deploy them on public roads in a very near future. Guaranteeing safe operation of these systems is crucial as they are planned to carry passengers and share the road with other vehicles and pedestrians. Yet, there is no agreed-upon approach on how and in what detail those systems should be tested. Different organizations have different testing approaches, and one common approach is to combine simulation-based testing with real-world driving.

One of the expectations from fully-automated vehicles is never to cause an accident. However, an automated vehicle may not be able to avoid all collisions, e.g., the collisions caused by other road occupants. Hence, it is important for the system designers to understand the boundary case scenarios where an autonomous vehicle can no longer avoid a collision. Besides safety, there are other expectations from automated vehicles such as comfortable driving and minimal fuel consumption. All safety and functional expectations from an automated driving system should be captured with a set of system requirements. It is challenging to create requirements that are unambiguous and usable for the design, testing, and evaluation of automated driving systems. Another challenge is to define useful metrics for assessing the testing quality because in general, it is impossible to test every possible scenario.

The goal of this dissertation is to formalize the theory for testing automated vehicles. Various methods for automatic test generation for automated-driving systems in simulation environments are presented and compared. The contributions presented in this dissertation include (i) new metrics that can be used to discover the boundary cases between safe and unsafe driving conditions, (ii) a new approach that combines combinatorial testing and optimization-guided test generation methods, (iii) approaches that utilize global optimization methods and random exploration to generate critical vehicle and pedestrian trajectories for testing purposes, (iv) a publicly-available simulation-based automated vehicle testing framework that enables application of the existing testing approaches in the literature, including the new approaches presented in this dissertation.

Lyric classification and generation are trending in topics in the machine learning community. Long Short-Term Networks (LSTMs) are effective tools for classifying and generating text. We explored their effectiveness in the generation and classification of lyrical data and proposed methods of evaluating their accuracy. We found that LSTM networks with dropout layers were effective at lyric classification. We also found that Word embedding LSTM networks were extremely effective at lyric generation.

In the field of machine learning, reinforcement learning stands out for its ability to explore approaches to complex, high dimensional problems that outperform even expert humans. For robotic locomotion tasks reinforcement learning provides an approach to solving them without the need for unique controllers. In this thesis, two reinforcement learning algorithms, Deep Deterministic Policy Gradient and Group Factor Policy Search are compared based upon their performance in the bipedal walking environment provided by OpenAI gym. These algorithms are evaluated on their performance in the environment and their sample efficiency.

Reinforcement learning (RL) is a powerful methodology for teaching autonomous agents complex behaviors and skills. A critical component in most RL algorithms is the reward function -- a mathematical function that provides numerical estimates for desirable and undesirable states. Typically, the reward function must be hand-designed by a human expert and, as a result, the scope of a robot's autonomy and ability to safely explore and learn in new and unforeseen environments is constrained by the specifics of the designed reward function. In this thesis, I design and implement a stateful collision anticipation model with powerful predictive capability based upon my research of sequential data modeling and modern recurrent neural networks. I also develop deep reinforcement learning methods whose rewards are generated by self-supervised training and intrinsic signals. The main objective is to work towards the development of resilient robots that can learn to anticipate and avoid damaging interactions by combining visual and proprioceptive cues from internal sensors. The introduced solutions are inspired by pain pathways in humans and animals, because such pathways are known to guide decision-making processes and promote self-preservation. A new "robot dodge ball' benchmark is introduced in order to test the validity of the developed algorithms in dynamic environments.