Stanford CS224R Deep Reinforcement Learning | Spring 2025 | Lecture 12: Multi-Task RL

The podcast discusses model-based reinforcement learning, including using learned models with synthetic data generation and determining when to use model-based reinforcement learning. It also covers multi-task imitation learning and reinforcement learning, including conditioning on tasks, goal-reaching tasks, and an approach called hindsight relabeling. The discussion includes planning with gradient-based or sampling-based optimization, updating models with collected data, and replanning to account for errors. The podcast further explores using learned models to learn a policy by augmenting collected data with a learned simulator, generating synthetic data, and updating policies using both real and generated data. Additionally, it addresses multi-task reinforcement learning, focusing on learning a generalist policy conditioned on the task, amortizing complexity across tasks, and leveraging shared structures between tasks, including identifying tasks and using task identifiers.