Reinforcement Learning (RL) has emerged as a powerful paradigm for addressing challenging decision-making and robotic control tasks. By leveraging the principles of trial-and-error learning, RL algorithms enable agents to learn optimal strategies through interactions with an environment.

However, despite the success of existing RL algorithms, their practical application in the real world still faces several challenges, such as sample inefficiency and the lack of interpretability arising from the reactive nature of RL policies. To tackle these challenges, we first propose an automated curriculum learning method for RL agents in dynamic environments. Next, we propose a hybrid approach combining reinforcement learning with a conventional search-based motion planner. Finally, we explore a novel offline reinforcement learning approach that uses latent diffusion for batch-constrained Q-learning. The proposed methods try to facilitate the adoption of RL in real-world settings, enabling intelligent decision-making and control in safety-critical domains.