Recent progress in robot learning has led to impressively capable manipulation systems. Much of the progress has come from scaling up human demonstrations; however, collecting such data through manual teleoperation is slow, costly, and hard to scale.

Physics-based Simulation offers a scalable, safe, and efficient alternative for generating large demonstration datasets. 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.

In this thesis, we tackle 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 and serve as a strong prior for downstream fine-tuning. Second, I will introduce the Generative Simulation pipeline for automatic generation of large-scale simulation datasets with minimal human efforts. 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.

Committee:
Zackory Erickson (co-chair)
David Held (co-chair)
Katerina Fragkiadaki
Chuang Gan (UMass Amherst and MIT-IBM Watson AI Lab)
Dieter Fox (University of Washington and Ai2)