Model-Based Reinforcement Learning with Learned World Models
Agents that learn a model of their environment and improve by planning or imagining inside it, rather than generating worlds for people.
Tier Grades
Tier A: Frontier AI labs
feasible
Google DeepMind develops both Dreamer and MuZero.
Tier B: Indonesian AI labs
stretch
DreamerV3 runs with one configuration across over 150 tasks, but compute is not stated in the abstract.
Tier C: Institution units such as ITB
feasible
A fixed configuration across over 150 tasks makes DreamerV3 a usable research baseline for university labs.
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
stretch
No industrial use is shown in the sources; use in a proof of concept would be exploratory.
Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.
Anchors
- ComputeDreamerV3: training compute not stated in the abstract; one configuration across over 150 tasks.
- DataDreamerV3: collects Minecraft diamonds from scratch without human data or curricula.
- Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): 4 authors on the DreamerV3 paper (arXiv 2301.04104); 12 on MuZero (arXiv 1911.08265).
- Time horizonMulti-year.
- Shelf-life signalsMuZero plans with a learned model without known rules, and DreamerV3 uses one configuration across over 150 tasks.
- Pipeline roleTier A publishes algorithms; Tier C reproduces them on simulated tasks; Tier B and Tier D have no observed role yet.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Links
R&D TopicModel-Based Reinforcement Learning with Learned World Models
Models
DreamerV3Google DeepMind, world modelMuZeroGoogle DeepMind, reinforcement learningDreamerGoogle DeepMind, world modelDreamer 4Google DeepMind, world model
DreamerV3Google DeepMind, world modelMuZeroGoogle DeepMind, reinforcement learningDreamerGoogle DeepMind, world modelDreamer 4Google DeepMind, world modelUse Casesnone recorded
Who Runs Itnone recorded
Papers
- World ModelsDavid Ha et al., arXiv preprint, 2018 founding
- Mastering Diverse Domains through World ModelsDanijar Hafner et al., arXiv preprint, 2023 state of the art
- TD-MPC2: Scalable, Robust World Models for Continuous ControlNicklas Hansen et al., ICLR 2024 (per arXiv comment), 2023 state of the art
- Dream to Control: Learning Behaviors by Latent ImaginationDanijar Hafner et al., arXiv preprint, 2019 state of the art
- Model-based Reinforcement Learning: A SurveyThomas M. Moerland et al., arXiv preprint, 2020 survey
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- arxiv.org/abs/2301.04104 https://arxiv.org/abs/2301.04104
- arxiv.org/abs/1911.08265 https://arxiv.org/abs/1911.08265
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Model-Based Reinforcement Learning with Learned World Models", https://aurion.semai.id/topics/model-based-rl-with-learned-world-models/, records of October 2026.