In this research, we focus on developing efficient spatiotemporal representation learning frameworks for robotics, particularly in tasks that require learning dynamics and control from partial observations. Robotic systems frequently face challenges such as high dimensional sensory inputs, partial observability, and computational constraints, which traditional methods struggle to address effectively. Existing approaches often fail to balance computational efficiency, accuracy, and robustness in dynamic environments. First, we introduce RoboKoop, a task-conditioned representation learning framework that leverages Koopman operator theory to linearize complex system dynamics into efficient latent spaces. This approach significantly improves the stability, robustness, and sample efficiency of control policies in robotics tasks. Second, we present STEMFold, a spatiotemporal modeling framework that utilizes dynamic graph-based attention mechanisms to predict trajectories in multi-agent systems, even in scenarios with unobservable agents. Finally, we develop AdaCred, a novel causal graph-based credit assignment mechanism that dynamically prioritizes task-relevant features, enhancing policy learning efficiency, particularly in offline and partially observable settings. Together, these contributions form a unified framework for building adaptive, scalable, and computationally efficient robotic systems capable of tackling real-world challenges. As the next step, we aim to extend these models toward real-time generalization and adaptation across diverse tasks. Specifically, we will focus on test-time adaptation frameworks that enable models to transfer across tasks with shared dynamics but differing objectives, as well as runtime monitoring mechanisms to ensure task success and mitigate failures.