danijar/dreamerv3

Mastering Diverse Domains through World Models

What it solves

It addresses the difficulty of applying reinforcement learning (RL) across different domains, which typically requires extensive expert knowledge and computational resources to tune hyperparameters. DreamerV3 provides a scalable, general-purpose RL algorithm that works across diverse tasks with a fixed set of hyperparameters.

How it works

DreamerV3 uses a world model to learn from experiences. It encodes sensory inputs into categorical representations and predicts future representations and rewards based on actions. The system then trains an actor-critic policy using "imagined trajectories" generated by this world model rather than relying solely on real-world interaction.

Who it’s for

Researchers and developers working in reinforcement learning and control tasks who want a robust, data-efficient agent that can scale with larger models and more gradient steps without needing task-specific tuning.

Highlights

  • Fixed Hyperparameters: Masters a wide range of domains without needing manual tuning.
  • Scalability: Performance and data efficiency consistently improve as model size and gradient steps increase.
  • World Model Approach: Learns a predictive model of the environment to train policies in imagination.
  • JAX-based: Implemented using JAX for efficient computation on GPU, CPU, or TPU.

Related

  • Project
    nicklashansen/tdmpc2

    A scalable, robust model-based reinforcement learning algorithm for continuous control that can handle multiple tasks across different domains and embodiments using a single agent.

  • Project
    nvidia-cosmos/cosmos-rl

    A scalable Reinforcement Learning framework specialized for Physical AI applications, providing a decoupled architecture for large-scale training workloads.

  • Project
    araffin/sbx

    A Jax-based proof-of-concept implementation of Stable-Baselines3, providing high-performance reinforcement learning algorithms with a consistent API.

  • Project
    PrimeIntellect-ai/prime-rl

    A framework for large-scale asynchronous reinforcement learning designed to train 1T+ parameter MoE models across 1000+ GPUs using FSDP2 and vLLM.

  • Project
    RLinf/RLinf

    A flexible and scalable reinforcement learning infrastructure designed to streamline the training of Embodied and Agentic AI across diverse hardware and simulators.