DreamerV3-Style RSSM World Model

What problem does it solve?

DreamerV3 RSSM provides a scalable, differentiable world-model that combines a deterministic BlockGRU memory with a stochastic unimix-enabled RSSM, enabling stable imagination rollouts and sample-efficient actor-critic training for complex control tasks.

Core Features & Use Cases

• Deteministic-stochastic state separation with BlockGRU memory and unimix-based sampling for robust sequence modeling.
• Imagination rollouts enabling actor-critic training without environment interaction, improving sample efficiency.
• KL balancing with free nats to prevent posterior collapse, plus support for multiple model sizes and configurations.
• Symlog-twohot prediction heads for rewards and continuation, providing scale-invariant learning across diverse domains.
• Self-contained templates, references, and assets to accelerate experimentation, validation, and iteration in research and development.

Quick Start

Clone this skill into your project, install PyTorch, and run the provided diagnostic script to initialize the RSSM and validate a minimal observe-imagine workflow.