Abstract: Robots hold the promise of assisting us in everyday environments, yet today’s systems 
remain limited in both dexterity (handling complex, contact-rich manipulation) and generalization 
(operating reliably in unstructured, changing scenes). The dominant approach in robot learning 
attempts to address this by scaling up datasets collected from human demonstrations. In this talk, 
I will argue that this approach faces fundamental limitations, and instead advocate for a shift 
toward learning from autonomous experience -- where robots actively explore, interact with the 
environment, and improve through their own trial and feedback. I will outline the key building blocks 
required to make this paradigm practical on real robots, and present my research contributions to 
each. Together, these components enable robots to acquire dexterous skills that generalize across 
objects, tasks, and environments.

Bio: Toru Lin is a PhD student at Berkeley AI Research (BAIR) Lab and a visiting researcher at Meta 
FAIR, advised by Jitendra Malik. Her research lies at the intersection of robotics and artificial 
intelligence, with a focus on learning dexterous and generalizable robot skills from autonomous 
experience. She received her BSc and MEng from MIT EECS, where she worked with Phillip Isola and 
Antonio Torralba, and previously studied at the University of Tokyo. She has interned at NVIDIA 
(GEAR group), DeepMind, Facebook, and Google. Her research is supported by the NSF Graduate 
Research Fellowship and the Berkeley Chancellor’s Fellowship