In 2026, world models are rapidly gaining traction in the field of embodied AI, with multiple companies launching related solutions, yet their technical paths and objectives vary significantly. Currently, embodied AI primarily relies on imitation learning, which shows clear limitations in complex environments. World models, however, can help AI establish environmental representations to aid decision-making. Lumiere Tech believes that robots also need to understand the causal chain between actions, consequences, and world changes. Therefore, they propose the 'Physics-Native Intelligence' technical route and have released the first model, Phi-WM 1.0 ActEffect. This model achieves precise observation, clear thinking, and stable operation through state-decoupled representation, temporal causality-driven mechanisms, and physical law constraints. During training, the model incorporates action causality into consideration. After training, the understanding of action consequences is distilled into the policy model, thereby reducing computational costs during deployment. Meanwhile, by leveraging multiple types of data, it enhances data efficiency and has achieved outstanding results in multiple benchmark tests. Lumiere Tech adheres to a software-hardware integrated development strategy. Its industrial-grade robot, Phi-Bot X1, has been validated and deployed in automotive factories, and it is currently building a system where the embodied brain, body, data, and scenarios evolve together. In the future, world models are expected to become a fundamental component of embodied AI, with evaluation criteria placing greater emphasis on performance in real-world scenarios.