Achieving long-horizon planning and decision-making capabilities, such as cooking a meal or cleaning a messy room, is essential for building intelligent robots. However, these tasks are challenging because they require selecting appropriate high-level actions, such as picking up a cup, and effectively planning how to execute those actions while satisfying physical and geometric constraints. We design task and motion planning algorithms to effectively address such challenges. Additionally, we are interested in exploring new problem classes by introducing practically relevant constraints and objectives.

We aim for robots to learn from past experiences and adapt their behaviors in a generalizable and robust manner to handle novel tasks. Recent advancements in deep learning and foundation models offer significant opportunities for robotics to embrace data-driven approaches by leveraging breakthroughs in computer vision and natural language processing. Robot learning encompasses, but is not limited to, the development of robot foundation models, the use of LLMs, and the learning of representations, symbols, skills, policies, and heuristics. In designing the learning pipeline, we often consider approaches such as few-shot learning, lifelong learning, and curriculum learning.

Robots are often deployed in environments where they coexist with other robots or humans, working collaboratively toward a common goal. Interactions with other robots involve challenges such as task decomposition, allocation, and multi-robot planning. In contrast, interactions with humans require inferring human intentions or behaviors and dynamically adapting the robot's strategy based on these inferences. Ad-hoc teamwork aims to address these challenges, enabling effective and seamless human-robot collaboration.