Parameterized Skill

Parameterized skills in robotics and AI aim to create reusable, adaptable actions that can be adjusted for various tasks and environments, improving efficiency and generalization compared to learning each task from scratch. Current research focuses on learning these skills from offline data using methods like vision-language models and reinforcement learning, often incorporating hierarchical structures and continuous parameterization for flexibility. This approach is significantly advancing robot learning, enabling more efficient task acquisition and adaptation in complex, real-world scenarios, and impacting fields like autonomous driving and human-robot collaboration. Furthermore, research is exploring how to best leverage these skills in conjunction with large language models and constraint satisfaction techniques for long-horizon planning.