Visual World Models in Embodied Intelligence: A Study on Physical Commonsense Embedding and Active Perception

Abstract

In 2026, embedded intelligence will enter the first year of mass production and application. At this time, the visual world model (VWM) has become the core to make agents understand the dynamic environment. However, when dealing with complex physical interaction, the existing VWM usually encounters the core contradiction between “visual fidelity” and “physical correction”. This is manifested in “physical illusion” and “long-term drift”. In order to solve this problem, this paper systematically studies the evolution of VWM, and focuses on the key challenge of “embedding physical common sense”. We propose an architecture scheme that combines explicit physical constraints with implicit neural representation. This enhances the model perception of physical laws such as gravity and collision, which is attributed to the differentiable physical loss function and the explicit geometric representation through 3D Gaussian Splicing (3DGS). In addition, this paper discusses the counterfactual reasoning mechanism based on causal intervention to improve the ability to predict the consequences of actions. It also builds a standardized evaluation benchmark, including physical rigor, visual quality and embedded task performance. This paper aims to provide a system reference for the construction of the next generation embedded intelligent system with “physical intuition”.