Abstract: In this talk, I will share my research for building generalist robots that can 
perform diverse manipulation tasks in unstructured, everyday environments. While recent 
progress in robot learning has come from scaling up human demonstrations, collecting 
such data through manual teleoperation is slow, costly, and hard to scale. My research 
takes a different approach by leveraging physics-based simulation as a scalable data
engine for robot learning. However, some core challenges limit the full potential of this 
approach: the heavy manual effort required to design simulation tasks and rewards, the 
gap between simulation and reality, and the difficulty of learning policies that generalize 
well when trained on large, diverse simulation datasets. My research tackles these 
challenges through scalable, generalizable, and adaptive robot learning. First, I will show 
how structured policy representations can enable simulation trained policies to achieve 
broad generalization in the real world. Second, I will introduce the Generative Simulation 
pipeline for automatic generation of large-scale simulation datasets. Finally, I will discuss 
some novel algorithms for efficient adaptation of simulation-trained policies to the real 
world. Together, these efforts move us toward robots that can learn broadly, adapt quickly, 
and assist people in real homes and workplaces.

BIO: Yufei Wang is a fifth-year PhD student at Carnegie Mellon University. His research 
interests span robot learning, embodied intelligence, and assistive robots. He has 
published over 20 papers at top robotics/machine learning conferences and journals such 
as NeurIPS, ICML, RSS, ICRA, CoRL, and RA-L. Yufei is supported by the Uber Presidential 
Fellowship and the Softbank & Arm Fellowship