
DreamerV3: Mastering Diverse Control Tasks Through World Models (Danijar Hafner et al.)
Danijar Hafner and colleagues present DreamerV3, which learns a world model and improves its behavior using imagined trajectories.
A single algorithm configuration supports training across more than 150 tasks, including games and continuous control.
The discussion follows the world model, actor, critic, and techniques that stabilize learning across different domains.
Dreamer combines discrete stochastic representations with a recurrent state; successful control depends on the information those representations preserve.
A comparison with V-JEPA 2 examines another route from predictive representations and robot interaction data to planning in new settings.
Author: Danijar Hafner, Jurgis Pasukonis, Jimmy Ba and Timothy Lillicrap
Source: Mastering Diverse Control Tasks Through World Models · Danijar Hafner, Jurgis Pasukonis, Jimmy Ba and Timothy Lillicrap
Series: AI Papers · Research deep dive
Explore: AI & Computer Science · AI Papers
Related episodes: PlaNet: Learning Latent Dynamics for Planning from Pixels (Danijar Hafner et al.) · Duo: A Reciprocal World Model for Full-Duplex Interaction (Hengyuan Cao et al.)
Primary paper: © The Authors 2025, CC BY 4.0. This episode is commentary. Companion reading: Mahmoud (Mido) Assran and colleagues, V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning (2025).
Flere episoder fra "Marginalia"



Gå ikke glip af nogen episoder af “Marginalia” - abonnér på podcasten med gratisapp GetPodcast.








