DreamerV3

DreamerV3 is the first model-based reinforcement learning algorithm to master diverse domains using fixed hyperparameters.

Developed by Danijar Hafner and Google DeepMind, DreamerV3 masters tasks across discrete and continuous action spaces without per-task tuning. It successfully trains on everything from the Atari 57 benchmark to Minecraft (collecting diamonds from scratch) and UR5 robot arm manipulation. By learning a world model from experience and planning in latent space, it achieves state-of-the-art performance while maintaining high data efficiency. This robustness makes it a premier choice for complex, multi-modal environments where manual hyperparameter optimization is impractical.

https://danijar.com/project/dreamerv3/