"Xieyue Intelligence" has closed its Angel+ round of financing worth hundreds of millions of RMB, with Linear Capital joining as a co-investor.

Recently, Xieyue Intelligence has completed a financing of several hundred million yuan in the Angel+ round. With the completion of this round of financing, Xieyue will continue to promote embodied foundational model training, the construction of computing power and data infrastructure, the expansion of the core team, as well as the R&D of home robot prototypes and scenario verification.

Founded in February 2026, Xieyue Intelligence is an embodied foundational model company that takes home as its first landing scenario. Xieyue is building a general embodied model oriented to the real physical world, driving the implementation of general home robots with self-developed models.

Xieyue was co-founded by Chen Wei, former Chief AI Scientist and head of the foundational model department at Li Auto and Zhang Xiao, former President of Product Line at Li Auto. Its core team members come from globally renowned technology, AI, robotics and intelligent vehicle enterprises, with capabilities covering five major fields: large model algorithms, AI Infra infrastructure, robot motion control, product and supply chain management. It is one of the top few teams in China with full-link experience in large-scale thousand-GPU large model training, digital and physical large model and agent development, and C-end product delivery from definition to mass production.

Xieyue believes that home is one of the most complex, highest-frequency and most long-term valuable real physical environments. Home tasks naturally integrate perception, understanding, planning, operation and interaction, covering both mental tasks and physical tasks, making it an ideal scenario to verify the capabilities of general embodied intelligence. Xieyue believes that home is the core training ground, verification ground and large-scale landing scenario for embodied foundational models. For the more long-term and ultimate goal of robots "entering every home", Xieyue proposes an embodied foundational model driven by the Duplex Reasoning paradigm.

In the past, traditional robot systems usually adopted a simplex or half-duplex process of "receiving instructions - executing actions - returning results", emphasizing correct action execution and task completion, while home scenarios need to respond to vague, changing and ongoing demands at all times. Duplex Reasoning aims to enable robots to maintain a "two-way channel" connected with humans in the process of perceptual understanding, reasoning planning and task execution, so that dialogue information can be interrupted and corrected at any time, and action goals can also be dynamically adjusted during the execution process, ultimately creating a general home action intelligence that is interruptible, correctable, takeoverable and continuously evolvable.

Home is not only a real physical world, but also people-centered. While executing digital skills and physical actions, Duplex Reasoning continuously receives human intervention, environmental feedback and safety signals. Instead of executing tasks unidirectionally, it continuously reasons and adjusts in interaction, action and feedback, realizing the integration of interaction and action.

The Duplex Reasoning proposed by Xieyue establishes a new paradigm for embodied foundational models: Interaction and action should be modeled in an integrated manner in data, model structure and training process. Taking VLA as the backbone, it connects the synergy with the prediction of the world model — language compresses vision, speech and environmental changes into transferable high-level semantics, driving the model to generalize across scenarios; the world model predicts the possible future outcomes of actions, providing pre-constraints for long-horizon tasks. Both point to one goal: to improve the success rate of robots' understanding, decision-making and action in open environments at a controllable computing power cost.

Xieyue believes that the core competitiveness of embodied intelligence does not lie in a larger model or more data volume, but in high-quality data and systematic infrastructure capabilities. Among them, training Infra and data Infra are two equally important core pillars — the training system determines the upper limit of model capabilities, and the data pipeline determines the implementation accuracy of the model. The company decomposes model capabilities into infrastructure problems that can be accumulated continuously:

First, training Infra: Build a reusable and scalable training system around pre-training, post-training and reinforcement learning to fully release the value of high-quality data;

Second, data Infra: Build a high-standard data pipeline around signal synchronization, task design and labeling quality, and prioritize high-representative, high-information-density premium data instead of blindly stacking data hours;

Third, establish a closed loop covering evaluation and simulation so that the progress of the model can be measured continuously and objectively.

Driven by both training Infra and data Infra, Xieyue will explore the Scaling Law of embodied intelligence at a more controllable computing power cost.

At present, Xieyue has initially completed the construction of self-developed Ego data collection equipment and data platform, and gradually built a layered data system around first-person perspective, non-prototype data and high-quality data related to prototypes. The company plans to run through the full capability from collection, cleaning, labeling to training within 2026, and continuously improve signal quality, task coverage and label effectiveness.

In terms of hardware, Xieyue adopts a phased strategy of "single prototype, full-stack closed loop". By controlling the consistency of prototype configuration, sensor selection and software and hardware interfaces, it first reduces the hardware variables in model R&D, and then imports the verified capabilities into the C-end-oriented home robot prototype. At the same time, Xieyue will continue to explore the Home-native product form, so that the robot can achieve a balance between spatial passability, interaction mode, safety boundary and home aesthetics.

In terms of commercialization, Xieyue will follow the pace of "verification first, then entering homes". Xieyue plans to first verify the model's cross-space generalization, task completion rate and unit economy in semi-structured scenarios such as hotels and nursing homes, and then gradually enter homes. High-frequency, long-horizon tasks with relatively clear evaluation standards such as laundry, storage and cleaning will become priority exploration directions.

At the current stage, what Xieyue is developing covers motion capability, generalization capability and personalized evolution capability, covering the closed loop of robot prototype, embodied foundational model and home self-evolution. Xieyue is traversing every factor that affects robots entering homes to build full-stack long-term capabilities — drive the embodied foundational model with Duplex Reasoning, drive the physical world data flywheel with Human-centric, build the robot safety system with Safety-first, and realize the natural integration into the real home environment with the Home-native dual-form prototype.

Regarding this round of financing, Chen Wei, Founder of Xieyue Intelligence, said: "Since Xieyue was founded half a year ago, we are increasingly convinced that building an embodied foundational model is a difficult but right thing to do. To build a model, in a sense, is to build solid ramparts and fight steady battles — we will build the full-link pipeline of computing power, data and evaluation solidly, and lay a solid foundation for the underlying infrastructure.

The current embodied intelligence field is far from convergence, with different data collection methods, different prototype forms, and great differences in the combination paths of models and hardware. This means that there is no end-to-end open-source model that can cover the whole link, and the model capability itself is the core competitiveness of embodied intelligence companies.

Xieyue adheres to the paradigm of large models to solve embodied problems, and has proved Xieyue's value to investors with phased achievements: continuous leading model performance, efficient and reliable task generalization, and the Duplex Reasoning paradigm has formed an iterative closed loop in home scenarios. We will dig deeper and wider the moat of home embodied intelligence step by step."

This article is from the WeChat official account "Linear Capital", published by 36Kr with authorization.