Dreamer 4 - DeepMind's new world model intelligent agent

What is Dreamer 4?

Dreamer 4, developed by DeepMind, is a novel intelligent agent that solves complex control tasks by training through visualization in a fast and accurate world model. In the game Minecraft, Dreamer 4 successfully obtained diamonds using only offline data, representing a significant breakthrough in the field. Its behavior is trained through reinforcement learning in a world model, eliminating the need for online interaction with the environment. This offers greater safety and efficiency in practical applications such as robotics, where online interaction can be risky and inefficient.

Dreamer 4's world model, based on an efficient Transformer architecture and a new shortcut forcing objective, enables real-time interactive inference on a single GPU. It can learn general action conditions from limited data and absorb most of the knowledge from large amounts of unlabeled video.

Main features of Dreamer 4

• Solve complex tasks through imagination trainingDreamer 4 uses reinforcement learning in its fast and accurate world model to solve complex control tasks without needing online interaction with the environment. This was demonstrated in the game Minecraft by successfully obtaining diamonds using only offline data, showcasing its powerful learning capabilities.
• High-efficiency real-time interactive reasoningBased on the efficient Transformer architecture and the new shortcut forcing objective, Dreamer 4's world model can achieve real-time interactive inference on a single GPU, making it more efficient and practical in real-world applications, especially in scenarios requiring fast response.
• Learning action conditions from a small amount of dataDreamer 4 can learn action conditions from a small amount of labeled action data and absorb most of the knowledge from a large amount of unlabeled video. This paves the way for learning general world knowledge from diverse unlabeled web videos in the future, reducing the reliance on large amounts of labeled data.
• Supports diverse tasks and generalization capabilitiesDreamer 4's world model can learn specific tasks and generalize to unseen scenarios and tasks. For example, in Minecraft, it can learn from action data that only contains the Overworld and generalize to unseen dimensions such as Nether and End.
• The potential to provide universal world knowledge learningDreamer 4's design lays the foundation for future research, including pre-training on general internet videos, integrating long-term memory, language understanding, and utilizing limited online correction data, thus driving the development of intelligent agents in a wider range of applications.

The technical principles of Dreamer 4

• World Model and Imagination TrainingDreamer 4 learns and optimizes policies by constructing a world model that simulates the dynamics of the environment and then training within that model using imagination. This training method allows agents to learn complex control tasks through simulated experience without needing to interact with the real environment online, improving learning efficiency and safety.
• High-efficiency Transformer architectureIt adopts the efficient Transformer architecture as the basis of the world model. The architecture has powerful parallel computing capabilities and the ability to process long sequence data, enabling the model to efficiently process complex video inputs and action sequences, providing agents with more accurate environmental predictions and decision support.
• Shortcut Forcing ObjectiveA new training objective—Shortcut Forcing Objective—is introduced, which forces the model to learn shorter inference paths during training, reducing error accumulation during generation, improving model stability and generation quality, and supporting fast interactive inference.
• Mask autoencoder and action condition learningThe tokenizer is trained using a mask autoencoder technique, enabling it to learn more robust visual representations from partially occluded images. Dreamer 4 can learn action conditions from a small amount of labeled action data and generalize them to unlabeled video data, extracting useful information from large-scale unlabeled data and expanding the model's application scope.
• Multi-task learning and strategy optimizationBy inserting task inputs into the world model, Dreamer 4 enables multi-task learning, allowing the agent to adjust its policies according to different task objectives. It utilizes reinforcement learning in imagination training to optimize policies, further improving the agent's performance in complex tasks and enabling it to better adapt to different environments and task requirements.

Dreamer 4 project address

• Project official websitehttps://danijar.com/project/dreamer4/
• arXiv technical paper: https://www.arxiv.org/pdf/2509.24527v1

Application scenarios of Dreamer 4

• Agent training in complex game environmentsIn complex game environments such as Minecraft, Dreamer 4 successfully obtained diamonds using only offline data, demonstrating its learning and decision-making capabilities in complex environments.
• RoboticsDreamer 4's world model enables real-time interactive reasoning on a single GPU, holding potential application value in robotics. Robots can be trained in a simulated environment without online interaction with the real world, improving training security and efficiency.
• Generalize to unseen scenariosDreamer 4 can learn action conditions from a small amount of labeled action data and generalize to unseen scenarios. It exhibits a degree of adaptability when handling new and unseen tasks, opening up possibilities for the application of intelligent agents in dynamic environments.
• Universal World Knowledge LearningDreamer 4 is designed to provide a foundation for learning general world knowledge from diverse, unlabeled online videos. It can be applied to scenarios requiring broad world knowledge, such as autonomous driving and intelligent surveillance, providing intelligent systems with a more comprehensive understanding of their environment.
• Multi-task learning and strategy optimizationDreamer 4 supports multi-task learning and can adjust its strategies according to different task objectives. It has advantages in scenarios that require handling multiple tasks, such as smart homes and smart factories, and can flexibly adjust to different task requirements.