World Model

The concept of the world model originates from the autonomous machine intelligence architecture proposed by Yann LeCun; its core idea is to enable AI to internally construct a simulator of the operating rules of the physical world in order to predict the consequences of actions.

David Ha and Jürgen Schmidhuber proposed the 'World Models' framework in 2018, using VAE and RNN to build internal representations of environments.

World models are essentially learned dynamical equations of an environment—given the current state and action, they predict the next state.

Traditional world models typically only model visual prediction, with information such as actions and depth serving as conditional inputs or post-processing results, making it difficult for the model to understand the physical coupling relationships among the three.

A world model can allow robots to rehearse the consequences of an action in a virtual environment without actually executing it.

Being able to predict the future and being able to infer the correct action are two entirely different things.

Mattick believes the capabilities demonstrated by world models are overestimated, because the tasks currently handled by world models are not difficult from a control perspective.

Yann LeCun is one of the three godfathers of deep learning and Meta's Chief AI Scientist; he argues that the current large language model approach cannot lead to true AGI, and considers the lack of a world model to be the core bottleneck.

LingBot-World is an interactive video diffusion world model with a parameter scale of 1.3B, making it lightweight among comparable products.

Teams such as World Labs and Decart are exploring the "world model" approach of having neural networks directly simulate physical laws and environmental feedback.

World model plus VLA is becoming the mainstream architectural consensus in the field of embodied intelligence.

The capabilities of robots on-site at WRC 2026 are mostly driven by VLA (Vision-Language-Action models) or World Models.

A 2023 MIT study trained a GPT model to predict the next move in Othello games and, through probing techniques, found that the model's internal representations encoded the complete board state.

Researchers such as Yann LeCun advocate building AI architectures based on world models, arguing that language modeling alone cannot achieve genuine understanding and safe decision-making.

Transformer models excel at capturing statistical correlations in text, but lack the ability to build world models and cannot internally simulate the causal chains and long-term consequences of physical world operations.

Dexforce uses a combination of a world model and a humanoid robot to accomplish general manipulation tasks such as brewing coffee.

Morphi Robot employs a hybrid architecture combining a world model and VLA, capable of supporting complete long-horizon tasks lasting up to 15 minutes.