Bridging State and Action: Towards Continual Reinforcement Learning - 2023 Fall Robotics Colloquium

Explore cutting-edge research in artificial intelligence and reinforcement learning through this 53-minute colloquium talk by Khimya Khetarpal from Google Deepmind. Delve into a framework that enables AI agents to represent and reason about their environment using affordances, which play a dual role in reducing action possibilities and facilitating efficient transition model learning. Discover an approach for building temporally abstract partial models using affordances to reason and make predictions across different time scales. Gain insights into the trade-offs between single-step models and temporally extended partial models, supported by theoretical guarantees. Learn how the lens of affordances offers benefits such as faster planning, improved sample efficiency, and robust generalization. Understand Khetarpal's research goals in developing AI systems that can efficiently represent world knowledge, plan with it, and adapt to changes over time through learning and interaction.