skills.rest
sovr610
sovr610COMMUNITY
Explore all skills by this author →
WebsiteGitHub

DreamerV3-Style RSSM World Model

Build a recurrent state space model with BlockGRU and unimix sampling.

Updated Feb 28, 2026
Search Tags:#pytorch#reinforcement-learning#world-model#dreamerv3#rssm#blockgru
One-click install
npx skills add https://github.com/sovr610/refffiy --skill dreamerv3-style-rssm-world-model
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: DreamerV3-Style RSSM World Model
Source: https://github.com/sovr610/refffiy/tree/main/brain-ai-dev/skills/dreamerv3-rssm
Command: npx skills add https://github.com/sovr610/refffiy --skill dreamerv3-style-rssm-world-model

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires torch, pytest, and includes scripts (resource) and references (resource) and assets (resource) components.

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.

Frequently Asked Questions about DreamerV3-Style RSSM World Model

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I build a recurrent state space model for reinforcement learning in PyTorch?▼

A recurrent state space model (RSSM) fuses deterministic memory with stochastic categorical states to enable stable imagination rollouts. This DreamerV3-style implementation uses a BlockGRU memory combined with unimix sampling for robust sequence modeling.

Why does posterior collapse happen in world models and how do I prevent it?▼

Posterior collapse in world models is prevented using KL balancing with free nats. This DreamerV3 RSSM implementation applies this technique to ensure the stochastic latent state retains meaningful information during actor-critic training.

How do I train an actor-critic model using imagination rollouts without environment interaction?▼

Actor-critic training without environment interaction is achieved through imagination rollouts generated by a differentiable world model. This RSSM supports imagination rollouts to improve sample efficiency by simulating environments internally.

What is symlog-twohot prediction and why is it used in world models?▼

Symlog-twohot prediction is a technique used in world models to ensure scale-invariant learning across diverse domains. This RSSM uses symlog-twohot prediction heads for rewards and continuation predictions to handle varying scales effectively.

Can I use a BlockGRU-based memory for deterministic state separation in PyTorch?▼

BlockGRU-based memory can be used for deterministic state separation in PyTorch. This RSSM implementation combines a BlockGRU deterministic memory with a stochastic unimix-enabled state to provide robust sequence modeling.

How do I validate gradient flow and numerical stability in an RSSM implementation?▼

Gradient flow and numerical stability in an RSSM implementation are validated using self-tests. This DreamerV3-style model includes self-tests to validate gradient flow, tensor shapes, and numerical stability across the observe-imagine workflow.

Related Skills

Explore More Software Engineering Skills→
MakisuoMakisuo
1.6k

clickhousectl-local-dev

Install and configure a local ClickHouse environment with clickhousectl.

Community
Intermediate
JoeyJoziahJoeyJoziah

AgentDB Learning Plugins

Generate AgentDB reinforcement learning plugins and train them from experience data.

Community
Advanced
fabri07fabri07

AgentDB Learning Plugins

Automate creation and training of reinforcement learning plugins for autonomous agents.

Community
Advanced
hung-phanhung-phan
1

rnn

Analyze recurrent neural network architectures for sequential data processing.

Community
Intermediate