UCLA

From manipulators to self-driving cars, training robots in the real world is often tedious, expensive, and results in hardware wear-and-tear. Due to the aforementioned reasons, the concept of transferring useful skills from "sim2real" has become an extremely attractive avenue for robotics researchers. Training purely (or partially) from simulation boasts numerous immense benefits such as allowing researchers to explore dangerous state spaces, learn faster than real-time, and even employ multiple agents to learn in parallel. Despite this, there currently exists a prominent "sim2real gap", where skills and/or models learned from within simulation transfer poorly to the real world due to environment misalignment. Given the scarcity of physically adequate models, this sim2real gap is especially prominent in contact-rich scenarios as well as problem spaces concerning deformables, whether it be the manipulation of deformable objects or soft robots themselves. In this thesis, I present a culmination of our previous works tackling two key sequential research areas: 1) development of efficient, physically accurate simulators for soft robots and structures and 2) full end-to-end sim2real solutions for robotic deformable material handling for tasks such as stiff sheet folding and deformable linear object deployment. Throughout these works, we showcase the immense benefits of developing solutions with physical insight in the areas of simulation, perception, and robotic manipulation.