Developing Robots that Autonomously Learn and Plan in the Real World

Humans plan and solve many tasks with ease. They can grow to perform incredible gymnastics, prove that black holes exist, and produce works of art, all starting from the same base learning system. While learning methods such as deep reinforcement learning have shown progress in simulated planning and control problems, they struggle to produce the same diverse, intelligent behaviour, especially in systems that interact in the real world (robots). This talk aims to discuss these limitations, provide methods to overcome them and enable agents capable of training autonomously to become learning and adapting systems that require little supervision while performing diverse tasks. The talk will present a series of works covering new, more robust Sim2Real methods, offline RL methods for longer planning tasks, and advances in generalization in planning to enable the creation of a single large policy to control all types of robots across diverse tasks.