In recent years, artificial intelligence for embodied agents such as robots has become an active area of research. For agents to understand the real world and act appropriately, it is important not only to understand spatial structures based on images and three-dimensional point clouds, but also to integrate diverse modalities such as vision, language, action, touch, force, and sound. 

 

In our laboratory, we work on recognition tasks such as object recognition, three-dimensional mapping, and graph network construction using machine learning, while also pursuing research on Embodied AI that connects these capabilities to robot navigation and robot manipulation. Furthermore, we aim to build spatial foundation models that can be continuously updated and developed through interaction with real-world environments, by teaching humanoid robots advanced human behaviors and task execution and enabling them to interpret such knowledge in relation to objects, people, and their surroundings. By combining supervised learning, reinforcement learning, inverse reinforcement learning, and learning-based control, we seek to establish foundational robotic technologies that can flexibly adapt to environmental changes and perform diverse real-world tasks.