Safe and Deployable Reinforcement Learning for Reason and Action

Explore the cutting-edge intersection of reinforcement learning and real-world AI deployment in this comprehensive conference talk by University of Wisconsin-Madison professor Josiah Hanna. Discover how AI agents can transcend static dataset limitations by learning decision-making through their own experiences, with reinforcement learning serving as the primary mechanism for trial-and-error goal achievement. Examine three critical aspects of scaling RL toward practical applications: first, learn about the challenges and innovative solutions for integrating reinforcement learning into complete robotic systems through competitive robot soccer research. Second, investigate methods for building confidence in AI agents before their decisions impact real-world scenarios, addressing crucial safety and reliability concerns. Third, delve into the revolutionary application of RL for training large language models to develop reasoning capabilities, revealing promising new directions for creating agents that simultaneously master both reasoning and action. Gain insights from Hanna's extensive research background spanning autonomous driving, robotics, and machine learning, as he presents strategies for making reinforcement learning more broadly applicable to real-world domains while maintaining safety and deployability standards.