PlaNet

Reproducing the Deep Planning Network (PlaNet) by Hafner et al..

PlaNet is a model-based RL agent that leverages its world model to plan in latent space.

Dependencies:

• PyTorch
• DeepMind Control Suite
• ffmpeg (optional: to render inline jupyter animations)
• OpenCV
• imagemagick (save matplotlib animations as gifs)

Checklist:

• Generate random seed episodes
• Implement experience replay
• Define transition model / RSSM
• Define reward model
• Define decoder (observation model)
• Define state encoder
• Implement planner
• Implement data collection using planner
• Define training objective