Learning World Models and Physical Agents

Explore how to develop world models and physical agents for robotics through this 58-minute conference talk by Sherry Yang from Montreal Robotics. Discover the challenges of training agents in physical environments where robot interactions incur high costs, unlike low-cost digital environments where superhuman performance has been achieved in games like AlphaGo. Learn about methods for creating world models from large-scale real-robot interaction data and understand how these models can efficiently evaluate real-robot policies, including testing scenarios with novel objects and distractors that fall outside the training distribution. Examine techniques for using world models to perform reinforcement learning and planning to enhance robot policies as physical agents. Gain insights from Yang's research at the intersection of machine learning, reinforcement learning, and generative modeling, with particular focus on applications in robotics and AI for science.