Abstract
Robot learning holds promise in alleviating several real world problems, by performing complex behaviors in complex environments. But what is the right way to train these robots? Our methods on self supervision shows encouraging results on several tasks like grasping objects, pushing objects and even flying drones. One key challenge with these methods is data efficiency. In this talk I will present our recent techniques on improving data efficiency using physical adversaries, multi task learning and curriculums.

Research Qualifier Committee:
Abhinav Gupta
Chris Atkeson
Martial Hebert
Xiaolong Wang