MiniMBRL is a minimal implementation of model-based reinforcement learning algorithms. The goal is to implement the most basic version of the algorithms to understand the core concepts and to build on top of them.
Here is the list of algorithms that I plan to implement:
- Dyna: An Integrated Architecture for Learning, Planning, and Reacting. Paper Notes: Dyna
- World Models
- PlaNet: Planning Network
- Dream to Control: Learning Behaviors by Latent Imagination(Dreamer V1)
- Mastering Atari with Discrete World Models(Dreamer V2)
- DayDreamer: World Models for Physical Robot Learning
- Planning to Explore via Self-Supervised World Models
- Neural Network Dynamics for Model-Based Deep Reinforcement Learning
- Deep Reinforcement Learning in a Handful of Trials using Probabilistic Dynamics Models
- Transformers are Sample-Efficient World Models
- The Benefits of Model-Based Generalization in Reinforcement Learning (ICML 2024)
- Facing Off World Model Backbones: RNNs, Transformers, and S4 (NeurIPS 2023)
- Curiosity-driven Exploration by Self-supervised Prediction
- Curious exploration via structured world models yields zero-shot object manipulation
- SENSEI: Semantic Exploration Guided by Foundation Models to Learn Versatile World Models
- Learning to Model the World with Language
- Language-Guided World Models: A Model-Based Approach to AI Control
- Install and activate a new python3.8 virtualenv:
virtualenv mbrl_venv --python=python3.8source mbrl_venv/bin/activatepip install "gymnasium[all]"
pip install mujoco