Xieyue Intelligence Raises Hundreds of Millions of Yuan in Angel+ Round for Home Robot Push

The round drew participation from Linear Capital, Junshan Capital, Hongyi Capital and Yinshan Capital, among other investors. Xieyue positions itself as an embodied foundation model company whose first deployment scenario is the home, building a general-purpose embodied model for the real physical world and using its self-developed models to drive general-purpose home robots.

The company was co-founded by Chen Wei, former AI chief scientist and foundation model division head at Li Auto, and Zhang Xiao, former product line president at Li Auto. Its core team comes from technology, AI, robotics and smart vehicle companies, spanning large-model algorithms, AI infrastructure, robot motion control, product and supply chain management, according to the report. It is among a small number of teams in China with experience in 10,000-GPU large-model training, digital and physical large-model and agent development, and end-to-end consumer product definition to mass production and delivery.

Xieyue argues the home is one of the most complex, high-frequency and long-term valuable real physical environments. Home tasks naturally combine perception, understanding, planning, manipulation and interaction, including both mental and physical tasks, making the home an ideal scenario for validating general embodied intelligence. The company believes the home is a core training ground, validation ground and scaled deployment scenario for embodied foundation models.

To pursue long-term robot deployment in homes, Xieyue has proposed an embodied foundation model driven by the Duplex Reasoning paradigm. Traditional robot systems often follow a simplex or half-duplex process of receiving commands, executing actions and returning results, with emphasis on correct action execution and task completion. In home scenarios, however, robots must respond to vague, changing and continuous needs. Duplex Reasoning is intended to keep a two-way channel between the robot and people during perception, understanding, reasoning, planning and task execution. Dialogue information can be interrupted and revised at any time, and action goals can be adjusted dynamically during execution. The aim is to create general home action intelligence that can be interrupted, corrected, taken over and continuously evolved.

Under Duplex Reasoning, the robot receives human intervention, environmental feedback and safety signals while performing digital skills and physical actions. It does not execute tasks in one direction but continuously reasons and adjusts through interaction, action and feedback, integrating interaction and action. The company says the paradigm establishes a new framework for embodied foundation models: interaction and action should be modeled together in data, model structure and training processes. With VLA as the backbone, it connects with world model predictions. Language compresses vision, speech and environmental changes into transferable high-level semantics to drive cross-scenario generalization; the world model predicts possible future results of actions to provide prior constraints for long-horizon tasks. Both aim to improve robots' understanding, decision-making and action success rates in open environments at controllable computing cost.

Xieyue says the core competitiveness of embodied intelligence is not larger models or more data volume, but high-quality data and systematic infrastructure capabilities. Training infrastructure and data infrastructure are two core pillars: the training system determines the upper limit of model capability, while the data pipeline determines the precision of deployment. The company breaks model capability into infrastructure problems that can be accumulated continuously. For training infrastructure, it is building a reusable and scalable training system around pre-training, post-training and reinforcement learning to release the value of high-quality data. For data infrastructure, it is building a high-standard data pipeline around signal synchronization, task design and annotation quality, prioritizing representative, information-dense data rather than blindly stacking data hours. It is also building a closed loop covering evaluation and simulation so that model progress can be measured continuously and objectively. With these two drivers, Xieyue aims to explore the scaling law of embodied intelligence at more controllable computing cost.

The company has initially completed its self-developed Ego data collection device and data platform, building a layered data system around first-person view, body-free data and body-related high-quality data. It plans to run through the full capability from collection, cleaning and annotation to training within 2026, while improving signal quality, task coverage and label effectiveness. On hardware, Xieyue has adopted a phased strategy of a single body and full-stack closed loop. By controlling consistency in body configuration, sensor selection and software-hardware interfaces, it first reduces hardware variables in model development, then imports validated capabilities into a consumer-facing home body. It will also explore Home-native product forms that balance spatial passability, interaction methods, safety boundaries and home aesthetics.

For commercialization, Xieyue will follow a rhythm of validating first, then entering the home. It plans to first validate cross-space generalization, task completion rate and unit economics in semi-structured scenarios such as hotels and nursing homes, before gradually entering homes. High-frequency, long-horizon tasks with relatively clear evaluation standards, such as laundry, storage and cleaning, will be priority directions. At the current stage, Xieyue is developing action capability, generalization capability and personalized evolution capability, covering the robot body, embodied foundation model and a home self-evolution loop. The company says it is working through every factor affecting robots' entry into homes and building full-stack long-term capabilities: using Duplex Reasoning to drive the embodied foundation model, Human-centric design to drive a physical-world data flywheel, Safety-first principles to build a robot safety system, and Home-native dual-form bodies to naturally integrate into real home environments.

Commenting on the financing, founder Chen Wei said: 'Thanks to the investors for their trust and support. In the half year since Xieyue was founded, we have become more convinced that building an embodied foundation model is difficult but right.' He added that building models is in some sense 'building hard camps and fighting dull battles'—solidly building the full-chain pipeline of computing, data and evaluation and laying a firm foundation. 'The embodied intelligence field is far from convergence, with different data collection methods, different body forms and very different paths for combining models and hardware. This means there is no end-to-end open-source model that can cover the entire chain. Model capability itself is the core competitiveness of an embodied intelligence company,' Chen said. He said Xieyue insists on using the large-model paradigm to solve embodied problems and has proven its value to investors with staged results: continuously leading model performance, efficient and reliable task generalization, and an iterative closed loop formed by the Duplex Reasoning paradigm in home scenarios. 'We will dig deeper and wider the moat of home embodied intelligence one shovel at a time,' he said.

Editor's Summary

Xieyue Intelligence has raised hundreds of millions of yuan in an angel-plus round less than a year after its founding, with backing from Linear Capital, Junshan Capital, Hongyi Capital and Yinshan Capital, among others. The company plans to use the capital for embodied foundation model training, computing and data infrastructure, team expansion and home robot development. Its Duplex Reasoning model and 'validate first, then enter homes' commercialization plan target semi-structured venues such as hotels and nursing homes before residential deployment.