Our research focuses on multimodal concept learning and reasoning, temporal dynamics modeling of humans and objects, and transfer learning of representation and skills to embodied agents. I believe multimodal learning is a pathway for computer vision to help language understanding, robotics, and social science.