Autonomous systems powered by Artificial Neural Networks (NNs) have shown remarkable capabilities in performing complex tasks that are difficult to formally specify. However, ensuring the safety, reliability, and trustworthiness of these NN-based systems remains a significant challenge, especially when they encounter inputs that fall outside the distribution of their training data. In robot learning applications, such as lower-leg prostheses, even well-trained policies can exhibit unsafe behaviors when faced with unforeseen or adversarial inputs, potentially leading to harmful outcomes. Addressing these safety concerns is crucial for the adoption and deployment of autonomous systems in real-world, safety-critical environments. To address these challenges, this dissertation presents a neural network repair framework aimed at enhancing safety in robot learning applications. First, a novel layer-wise repair method utilizing Mixed-Integer Quadratic Programming (MIQP) is introduced that enables targeted adjustments to specific layers of a neural network to satisfy predefined safety constraints without altering the network’s structure. Second, the practical effectiveness of the proposed methods is demonstrated through extensive experiments on safety-critical assistive devices, particularly lower-leg prostheses, to ensure the generation of safe and reliable neural policies. Third, the integration of predictive models is explored to enforce implicit safety constraints, allowing for anticipation and mitigation of unsafe behaviors through a two-step supervised learning approach that combines behavioral cloning with neural network repair. By addressing these areas, this dissertation advances the state-of-the-art in neural network repair for robot learning. The outcome of this work promotes the development of robust and secure autonomous systems capable of operating safely in unpredictable and dynamic real-world environments.

Human in the loop sequential decision making such as Reinforcement Learning from Human Feedback (RLHF) or behavior synthesis leverages human feedback to the AI system for multifaceted purposes. This dissertation will interpret the use cases of human feedback such as realizing individual human preferences, obtaining general common-sense human guidance and guidance for domain and dynamics information. While individual preferences are only known to the human in the loop, this dissertation shows that guidance information can be obtained from alternate, automated sources to mitigate the use of humans as a crutch. Specifically, RLHF on tacit tasks such as robot arm manipulation and locomotion suffer from high feedback complexity. A large portion of human-AI interaction budget is used by the AI agent in discovering guidance information rather than user preferences, essentially disrespecting their efforts. Similarly, for task-planning with human in the loop, a major challenge is acquiring common-sense user preferences on the agent behaviors. This dissertation proposes ways of obtaining priors (or background knowledge) to support guidance information needed by the AI agents thereby reducing the burden on the human in the loop. For RLHF in tacit tasks, the research minimizes unnecessary interaction by observing that a large budget of human feedback mostly informs the AI agent about domain structure information (such as temporal relationship between states) or the fact that human feedback is typically conditioned on a few key states. The thesis builds guidance priors based on these observations which provide effective means of reducing the interaction burden on the human in the loop. For symbolic task planning, the research explores the reliability of Large Language Models to act as a preference proxy for common-sensical guidance information. Specifically, we investigate along reasoning abilities in performing preferred plan detection and its brittleness in operationalizing human advice to generate plans. We extend this argument to approximate retrieval abilities for teasing out desirable domain models such as task reduction schemas for LLMs, which are useful in AI planning.

Robots are increasingly integrated into our daily routines, yet their current tasks are mostly specific and pre-programmed, lacking flexibility and scalability. In anticipation of a future where robots will handle diverse household chores—from basic tasks like picking and placing items to more complex activities such as cooking—there's a critical need for them to master long-term planning and motion challenges. Current methods addressing this demand typically rely on manually crafted abstractions and expert-guided task planning. This work deals with a novel approach - developing strategies to learn relational abstractions directly from raw trajectory data. These learned abstractions are then used for inventing symbolic vocabularies and action models. The learnt action models are then used to solve complex long-horizon task and motion planning problems, which are not seen in the training demonstrations. The results show that the approach discussed is robust and capable of learning the model just from few demonstrations. Additionally, this work also discusses an interactive AI platform aimed at making advanced robot planning accessible to users without extensive computer science backgrounds. Such platforms play a crucial role as AI and robotics increasingly intertwine with everyday life, offering intuitive interfaces that teach users the basics of robot planning.

While the growing prevalence of robots in industry and daily life necessitatesknowing how to operate them safely and effectively, the steep learning curve of programming languages and formal AI education is a barrier for most beginner users. This thesis presents an interactive platform which leverages a block based programming interface with natural language instructions to teach robotics programming to novice users. An integrated robot simulator allows users to view the execution of their high-level plan, with the hierarchical low level planning abstracted away from them. Users are provided human-understandable explanations of their planning failures and hints using LLMs to enhance the learning process. The results obtained from a user study conducted with students having minimal programming experience show that JEDAI-Ed is successful in teaching robotic planning to users, as well as increasing their curiosity about AI in general.

Recent advances in Artificial Intelligence (AI) have brought AI closer to laypeople than ever before. This leads to a pervasive problem: how would a user ascertain whether an AI system will be safe, reliable, or useful in a given situation? This problem becomes particularly challenging when it is considered that most autonomous systems are not designed by their users; the internal software of these systems may be unavailable or difficult to understand; and the functionality of these systems may even change from initial specifications as a result of learning. To overcome these challenges, this dissertation proposes a paradigm for third-party autonomous assessment of black-box taskable AI systems. The four main desiderata of such assessment systems are: (i) interpretability: generating a description of the AI system's functionality in a language that the target user can understand; (ii) correctness: ensuring that the description of AI system's working is accurate; (iii) generalizability creating a solution approach that works well for different types of AI systems; and (iv) minimal requirements: creating an assessment system that does not place complex requirements on AI systems to support the third-party assessment, otherwise the manufacturers of AI system's might not support such an assessment. To satisfy these properties, this dissertation presents algorithms and requirements that would enable user-aligned autonomous assessment that helps the user understand the limits of a black-box AI system's safe operability. This dissertation proposes a personalized AI assessment module that discovers the high-level ``capabilities'' of an AI system with arbitrary internal planning algorithms/policies and learns an accurate symbolic description of these capabilities in terms of concepts that a user understands. Furthermore, the dissertation includes the associated theoretical results and the empirical evaluations. The results show that (i) a primitive query-response interface can enable the development of autonomous assessment modules that can derive a causally accurate user-interpretable model of the system's capabilities efficiently, and (ii) such descriptions are easier to understand and reason with for the users than the agent's primitive actions.

In today's world, robotic technology has become increasingly prevalent across various fields such as manufacturing, warehouses, delivery, and household applications. Planning is crucial for robots to solve various tasks in such difficult domains. However, most robots rely heavily on humans for world models that enable planning. Consequently, it is not only expensive to create such world models, as it requires human experts who understand the domain as well as robot limitations, these models may also be biased by human embodiment, which can be limiting for robots whose kinematics are not human-like. This thesis answers the fundamental question: Can we learn such world models automatically? This research shows that we can learn complex world models directly from unannotated and unlabeled demonstrations containing only the configurations of the robot and the objects in the environment. The core contributions of this thesis are the first known approaches for i) task and motion planning that explicitly handle stochasticity, ii) automatically inventing neuro-symbolic state and action abstractions for deterministic and stochastic motion planning, and iii) automatically inventing relational and interpretable world models in the form of symbolic predicates and actions. This thesis also presents a thorough and rigorous empirical experimentation. With experiments in both simulated and real-world settings, this thesis has demonstrated the efficacy and robustness of automatically learned world models in overcoming challenges, generalizing beyond situations encountered during training.

Reinforcement Learning (RL) presents a diverse and expansive collection of approaches that enable systems to learn and adapt through interaction with their environments. However, the widespread deployment of RL in real-world applications is hindered by challenges related to sample efficiency and the interpretability of decision-making processes. This thesis addresses the critical challenges of sample efficiency and interpretability in reinforcement learning (RL), which are pivotal for advancing RL applications in complex, real-world scenarios.This work first presents a novel approach for learning dynamic abstract representations for continuous or parameterized state and action spaces. Empirical evaluations show that the proposed approach achieves a higher sample efficiency and beat state- of-the-art Deep-RL methods. Second, it presents a new approach HOPL for Transfer Reinforcement Learning (RL) for Stochastic Shortest Path (SSP) problems in factored domains with unknown transition functions. This approach continually learns transferable, generalizable knowledge in the form of symbolically represented options and integrates search techniques with RL to solve new problems by efficiently composing the learned options. The empirical results show that the approach achieves superior sample efficiency as compared to SOTA methods for transfering learned knowledge.

Robot motion and control remains a complex problem both in general and inthe field of machine learning (ML). Without ML approaches, robot controllers are typically designed manually, which can take considerable time, generally requiring accounting for a range of edge cases and often producing models highly constrained to specific tasks. ML can decrease the time it takes to create a model while simultaneously allowing it to operate on a broader range of tasks. The utilization of neural networks to learn from demonstration is, in particular, an approach with growing popularity due to its potential to quickly fit the parameters of a model to mimic training data. Many such neural networks, especially in the realm of transformer-based architectures, act more as planners, taking in an initial context and then generating a sequence from that context one step at a time. Others hybridize the approach, predicting a latent plan and conditioning immediate actions on that plan. Such approaches may limit a model’s ability to interact with a dynamic environment, needing to replan to fully update its understanding of the environmental context. In this thesis, Language-commanded Scene-aware Action Response (LanSAR) is proposed as a reactive transformer-based neural network that makes immediate decisions based on previous actions and environmental changes. Its actions are further conditioned on a language command, serving as a control mechanism while also narrowing the distribution of possible actions around this command. It is shown that LanSAR successfully learns a strong representation of multimodal visual and spatial input, and learns reasonable motions in relation to most language commands. It is also shown that LanSAR can struggle with both the accuracy of motions and understanding the specific semantics of language commands

The rise in popularity of applications and services that charge for access to proprietary trained models has led to increased interest in the robustness of these models and the security of the environments in which inference is conducted. State-of-the-art attacks extract models and generate adversarial examples by inferring relationships between a model’s input and output. Popular variants of these attacks have been shown to be deterred by countermeasures that poison predicted class distributions and mask class boundary gradients. Neural networks are also vulnerable to timing side-channel attacks. This work builds on top of Subneural, an attack framework that uses floating point timing side channels to extract neural structures. Novel applications of addition timing side channels are introduced, allowing the signs and arrangements of leaked parameters to be discerned more efficiently. Addition timing is also used to leak network biases, making the framework applicable to a wider range of targets. The enhanced framework is shown to be effective against models protected by prediction poisoning and gradient masking adversarial countermeasures and to be competitive with adaptive black box adversarial attacks against stateful defenses. Mitigations necessary to protect against floating-point timing side-channel attacks are also presented.

This dissertation introduces and examines Soft Curved Reconfigurable Anisotropic Mechanisms (SCRAMs) as a solution to address actuation, manufacturing, and modeling challenges in the field of soft robotics, with the aim of facilitating the broader implementation of soft robots in various industries. SCRAM systems utilize the curved geometry of thin elastic structures to tackle these challenges in soft robots. SCRAM devices can modify their dynamic behavior by incorporating reconfigurable anisotropic stiffness, thereby enabling tailored locomotion patterns for specific tasks. This approach simplifies the actuation of robots, resulting in lighter, more flexible, cost-effective, and safer soft robotic systems. This dissertation demonstrates the potential of SCRAM devices through several case studies. These studies investigate virtual joints and shape change propagation in tubes, as well as anisotropic dynamic behavior in vibrational soft twisted beams, effectively demonstrating interesting locomotion patterns that are achievable using simple actuation mechanisms. The dissertation also addresses modeling and simulation challenges by introducing a reduced-order modeling approach. This approach enables fast and accurate simulations of soft robots and is compatible with existing rigid body simulators. Additionally, this dissertation investigates the prototyping processes of SCRAM devices and offers a comprehensive framework for the development of these devices. Overall, this dissertation demonstrates the potential of SCRAM devices to overcome actuation, modeling, and manufacturing challenges in soft robotics. The innovative concepts and approaches presented have implications for various industries that require cost-effective, adaptable, and safe robotic systems. SCRAM devices pave the way for the widespread application of soft robots in diverse domains.