Real-World Reinforcement Learning for Mobile Robots
Mobile robots stand poised to transform our world, from food and package delivery to infrastructure maintenance and search-and-rescue. However, fully autonomous deployment of mobile robots remains a significant challenge. In this talk, I will first present a high-level overview of tools available to tackle this challenge, the pros and cons of each tool, and why real-world reinforcement learning is an appropriate tool for learning decision making systems for mobile robots. Next, I will discuss the criteria we care about when choosing and developing reinforcement learning algorithms---performance, sample-efficiency, safety, and supervision cost---and discuss which aspects are the bottleneck for mobile robot applications. Finally, I will present our work that seeks to address some of these bottlenecks, and which enabled a small-scale RC car to learn to avoid collisions in the real-world completely from scratch with only 4 hours of data and no a priori knowledge.
Gregory Kahn is a PhD student at UC Berkeley working with Professor Pieter Abbeel and Professor Sergey Levine in the Berkeley Artificial Intelligence Research (BAIR) Lab. Greg's main research goal is to develop algorithms that enable robots to operate in the real-world.