From Narrow Robots to General Robots
Despite the accelerating progress in robotics, robots today remain relatively narrow in their capabilities. To have robots that can work seamlessly with humans, I will advocate building "generalist robots" that are good at multiple tasks, in various complex environments. My research studies how to build generalist robots by learning to model the world. I will show that current ideas on building generalist robots have produced powerful results such as robot face that learns to mimic human facial expressions with a self-image, robots that play hide-and-seek by predicting the opponent's visual perspective, and algorithms that can distill compact physical knowledge from video recordings of multiple dynamical systems. Future directions will enable robots to have strong flexible and adaptable behaviors, rich perception systems from multiple modalities and novel scientific discovery for dynamical system modeling and control.