DreamerV3
Mastering Diverse Domains through World Models
...The system works by building an internal model of the environment and then using that model to simulate possible future outcomes of actions, allowing the agent to learn from imagined experiences rather than only from real interactions. This approach enables the algorithm to efficiently learn policies for decision-making tasks that would otherwise require enormous amounts of data or computational resources. DreamerV3 was designed as a general reinforcement learning framework that can solve diverse tasks using the same configuration of hyperparameters across many environments. In research demonstrations, the algorithm has been shown to perform strongly across more than one hundred control tasks and complex simulated environments.