This project investigates algorithmic and system-level challenges in federated learning (FL) for resource-constrained and strategically heterogeneous environments. FL is a distributed learning paradigm that enables collaborative model training across decentralized clients without centralizing raw data, making it well-suited for privacy-sensitive and communication-limited settings such as wireless networks. Despite its promise, FL faces fundamental challenges related to client heterogeneity, unreliable participation, and misaligned incentives that limit its practical deployment. This project addresses these challenges along three interconnected research lines. The first focuses on FL under energy harvesting constraints, where client devices rely on ambient energy sources and cannot guarantee continuous participation. The goal is to design pipelined training frameworks that accommodate intermittent availability without stalling global model convergence. The second research line develops principled client scheduling strategies based on information-theoretic notions of model staleness, aiming to improve convergence efficiency in settings where clients contribute updates at irregular intervals. The third line examines federated learning through the lens of mechanism design and economic theory, studying how incentive structures shape client behavior and affect learning outcomes, particularly in the presence of low-quality or strategically acting participants. Across all three lines, the core research objectives are to design algorithms with theoretical convergence guarantees, validate them through large-scale simulation experiments, and produce open-source implementations. Project outcomes will include novel federated learning algorithms, convergence analyses, and comprehensive experimental results targeting publication in leading venues in machine learning and communications.