UC Berkeley

Reinforcement learning has been a powerful tool for building continuously improving systems in domains like video games and animated character control, but has proven relatively more challenging to apply to problems in real world robotics. In this talk, I will argue that this challenge can be attributed to a mismatch in assumptions between typical RL algorithms and what the real world actually provides, making data collection and utilization difficult. In this talk, I will discuss how to build algorithms and systems to bridge these assumptions and allow robotic learning systems to operate under the assumptions of the real world - under realistic and practical assumptions on non-determinism, uncertainty and human supervision. In particular, I will describe how we can develop algorithms to ensure easily scalable supervision from humans, perform safe, directed exploration in practical time scales and enable uninterrupted autonomous data collection at scale. I will show how these techniques can be applied to real world robotic systems. Lastly, I will provide some perspectives on how this opens the door towards future deployment of robots into unstructured human-centric environments such as our homes, hospitals, and shopping malls.