TOWARDS KNOWLEDGE-DRIVEN AI FOR LARGE-SCALE OPTIMIZATION AND ROBOT LEARNING

In the pursuit of advancing artificial intelligence (AI) capabilities for addressing complex real-world challenges, it is imperative to integrate knowledge-driven approaches into AI systems while also leveraging AI to enhance knowledge representation and reasoning. This thesis explores Knowledge-driven Artificial Intelligence for Large-Scale Optimization and Robot Learning, addressing critical gaps in existing methodologies by unifying physics-based insights, learning-based strategies, and combinatorial optimization techniques.This thesis is composed of interrelated research directions. First, it investigates combinatorial optimization problems, which remain computationally intractable for classical algorithms as problem sizes scale. Focusing on the canonical Traveling Salesman Problem (TSP), a novel solution framework called Neuro-Ising is introduced, combining graph neural networks with localized Ising models. This hybrid approach demonstrates promising efficacy in solving large-scale TSP instances by leveraging the strengths of both neural approximations and physics-inspired energy minimization techniques. Second, the study explores dynamic optimization in autonomous systems through Model Predictive Control (MPC), a widely used paradigm in robotic motion planning and control. A Robust Adaptive MPC scheme (RAMP-Net) is proposed, integrating Physics-Informed Neural Networks (PINNs) to model robot dynamics and external disturbances such as wind effects, frictional forces, and actuator uncertainties. This approach addresses the limitations of conventional MPC, which often relies on simplified or inaccurate system models, and outperforms existing regression-based learning MPC methods in terms of tracking error. Third, the research extends into neuromorphic energy-efficient robot navigation, an area critical for real-time autonomous decision-making. A novel event-based physics-driven neuromorphic planner (EV-Planner) is presented, leveraging spiking neural networks (SNNs) and event-based vision in combination with depth sensing. This approach enables efficient and reactive motion planning for autonomous drones navigating dynamic environments, demonstrated through a scenario where a drone must fly through a moving ring while avoiding obstacles, along with real-world demonstration of the proposed algorithm. The thesis culminates with an Adaptive Safety Margin Algorithm (ASMA) for vision-language-based navigation, integrating Contrastive Language-Image Pretraining (CLIP) with Control Barrier Functions (CBFs) to enable constraint-aware, language-conditioned navigation. This approach refines AI-driven perception and control by incorporating semantic reasoning into real-time safety constraints, pushing the boundaries of human-intelligible robotic decision-making. These contributions advance the paradigm of knowledge-driven AI, offering scalable solutions for optimization and robotics that blend classical mathematical rigor with modern AI-driven adaptability.