Former Huawei AI Leaders Bet on a Physics Foundation Model for Embodied AI

Introduction

Large language models have demonstrated that broad capabilities can emerge from pretraining on massive text corpora. Robotics has followed a different path: systems are commonly optimized for a particular body, environment, and task, with new data collected whenever the deployment setting changes. Xirang Kaiwu, founded by former Huawei cloud large-model executive Li Yin and former Huawei multimodal scientist Zhang Hanwang, is attempting to build a shared foundation for the physical world. The company calls the proposed system a Large Physics Model, or LPM.

Key points

• Funding and organization: Xirang Kaiwu says it has completed several hundred million yuan in seed and angel financing, with a reported valuation of roughly $500 million. Its organization is divided into model, data, and systems-infrastructure functions.
• The problem: An LPM would need to understand not only what appears in an image, but also how objects move, how actions change the environment, and why a particular outcome occurred. The intended result is better transfer across robot bodies, materials, lighting conditions, and tasks.
• Data and architecture: The planned data mix includes internet videos of physical interactions, first-person data, demonstrations, simulation, public robot trajectories, and feedback from real machines. The main architecture is described as a unified autoregressive Transformer, with diffusion used when continuous future states or videos need to be rendered.
• Three capability layers: L0 is intended as a general physical foundation model; L1 would specialize it for industrial or other domains; L2 would convert those capabilities into deployable robot applications.

Why a physics foundation model?

Embodied intelligence must act in environments that evolve continuously. Consider clearing a table: a robot has to identify objects, choose an order, plan movements, and revise the plan after every action changes the scene. As the number of decisions grows, the space of possible action sequences expands rapidly. Demonstrations collected for one narrow task cannot realistically cover all of those possibilities.

Xirang’s proposed answer is to model the current state, the action, and the next state within one ongoing predictive process. The system chooses an action from the present state, observes the result, and continues reasoning from the updated state instead of enumerating every possible future in advance. Shared tokens, reinforcement learning, data pipelines, evaluation, and large-scale infrastructure are presented as parts of the same training loop.

Significance and open questions

The commercial promise is capability reuse: moving robotics away from “one deployment, one retraining” toward a model that carries physical knowledge between applications. If physical intelligence exhibits a scaling effect comparable to language modeling, larger datasets and models could produce more reliable generalization. The company says its model ranked first on trajectory accuracy and second on physical adherence in relevant tracks of World Arena 2.0, and that performance has shown an early upward trend as training scale increased. These are company-provided disclosures and still require independent testing.

The obstacles are substantial. Statistical correlations in video do not automatically become causal, executable knowledge; simulation results may fail to transfer to real hardware. Data quality, long-horizon planning, training cost, safety, and cross-body generalization will determine whether LPM becomes more than an appealing concept. The planned XIRRA v0.1 release may offer an early checkpoint for judging the approach.

Source: QbitAI