A living map of model-based RL, from Dyna to diffusion

What it does

This repository is a curated bibliography of model-based reinforcement learning research. It collects papers from foundational work like Sutton’s Dyna architecture and PILCO through the latest ICLR and ICML 2026 submissions, organizing them by conference and year. The list also includes a visual taxonomy, plus links to tutorials and reference codebases.

The interesting bit

The maintainers candidly admit that algorithm modularity breaks tree structures, yet they still provide a useful “Learn the Model” versus “Given the Model” split. It is rare for an awesome-list to openly confess the limits of its own taxonomy while still mapping the territory from 1991 to 2026.

Key highlights

• Chronological coverage spanning classic MBRL (Sutton 1991, Levine 2014) to contemporary diffusion world models and robot policy evaluation.
• Visual taxonomy diagram (mbrl-taxonomy.png) categorizing algorithms by model-learning versus model-planning approaches.
• Each paper entry includes authors, a one-line “Key” insight, and the experimental environments used.
• Sections for tutorials and codebases beyond pure paper listings.

Caveats

• The taxonomy is explicitly labeled “non-exhaustive” and the authors note tree structures poorly capture algorithm modularity.
• Entries are bibliographic, not critical reviews; there is little synthesis or comparison between methods.
• As a standard awesome-list, it is primarily an index rather than a runnable framework or reproduction suite.

Verdict

Worth bookmarking if you are an RL researcher or engineer trying to trace the lineage of world models, value expansion, and model-based policy optimization. If you need a unified training framework or hands-on tutorials, look elsewhere.