DreamerV3
A latent-space world model that performs predictions within a compressed latent representation space. It exhibits high computational efficiency and strong sample efficiency, and is applicable to a variety of robotic tasks.
Core Facts
Timeline (last 90 days)
DreamerV3 is a world-model-based reinforcement learning algorithm proposed by DeepMind that learned to mine diamonds from scratch in Minecraft, demonstrating the generality of a single hyperparameter configuration across many task types.
Works such as DreamerV3 and IRIS have advanced the world model framework toward higher-dimensional visual input tasks.
Algorithms such as DreamerV3 have demonstrated that AI agents equipped with world models surpass traditional methods in sample efficiency and generalization capability.
DreamerV3 was proposed by Danijar Hafner et al. in 2023, and was the first to achieve strong performance with a unified algorithm across more than 150 different tasks.
All Facts (5)
Institutions such as DeepMind and Meta AI have produced representative world model works including DreamerV3 and JEPA
65%UnverifiedDreamerV3 was proposed by Danijar Hafner et al. in 2023, and was the first to achieve strong performance with a unified algorithm across more than 150 different tasks.
70%UnverifiedDreamerV3 is a world-model-based reinforcement learning algorithm proposed by DeepMind that learned to mine diamonds from scratch in Minecraft, demonstrating the generality of a single hyperparameter configuration across many task types.
50%UnverifiedWorks such as DreamerV3 and IRIS have advanced the world model framework toward higher-dimensional visual input tasks.
50%UnverifiedAlgorithms such as DreamerV3 have demonstrated that AI agents equipped with world models surpass traditional methods in sample efficiency and generalization capability.
50%