X Square Robot Secures Hundreds of Millions in Dual Rounds Led by Hua Capital and Meituan, Advances China's First End-to-End Proprietary General Embodied Intelligence Model

On May 12, embodied AI company “Variable Robotics” announced the completion of two funding rounds totaling hundreds of millions of yuan: a Pre-A+++ round and an A round. The Pre-A+++ round was led by Huaying Capital, with participation from Yunqi Partners and GF Xinde. The A round was led by Meituan Strategic Investment, with participation from Meituan Longzhu. The proceeds from both rounds will be used to accelerate the synchronized iteration of its fully self-developed end-to-end general-purpose embodied AI large model and robot hardware, as well as to foster collaboration and deployment of intelligent solutions across multiple future application scenarios.

X Square Robot, founded in December 2023, focuses on the research and development of “general-purpose embodied large models,” leveraging real-world data as its primary source to build general-purpose robots with fine manipulation capabilities. To date, X Square Robot has completed seven rounds of financing, with cumulative funding exceeding RMB 1 billion.

The ultimate goal of general-purpose robots is to autonomously perform tasks through interaction, perception, and action, much like humans, while possessing efficient generalization and transfer capabilities. The core to achieving this objective lies in large foundation models for general embodied AI in robotics. Overseas, tech companies such as Skild AI, Google DeepMind, and Physical Intelligence (PI) are actively positioning themselves in this field.

Embodied AI can be broadly categorized into the “brain” (cognition and decision-making) and the “cerebellum” (motor control). Currently, Chinese companies are exploring diverse development pathways: some focus on the “brain,” enhancing robots’ language understanding and planning capabilities; others concentrate on the “cerebellum,” optimizing motor control functions such as locomotion and grasping. Some enterprises have adopted an end-to-end approach that integrates both the “brain” and “cerebellum,” a strategy also pursued by leading international tech companies such as Physical Intelligence (PI) and Skild AI.

From its inception, ZiBianLiang has pursued the route of “unified cerebellar-cerebral end-to-end large models.” As the first company in China to adopt this approach for developing general-purpose embodied AI large models, ZiBianLiang Robotics is committed to building an integrated system architecture that unifies perception, planning, and control. The company’s independently developed “WALL-A” model, part of its “Great Wall” series of operational large models, features multimodal information fusion capabilities. It can integrate diverse types of perceptual data, natural language instructions, and motion control signals, thereby achieving end-to-end mapping from input to output.

Leveraging large-scale general knowledge pre-training and multi-task learning mechanisms, the “WALL-A” model has demonstrated zero-shot generalization capabilities in certain unseen new task scenarios—zero-shot generalization across a wide range of scenarios is one of the key hallmarks of achieving general-purpose robots.

Meanwhile, Invariable Robotics has established a model-driven data closed-loop system. The company has independently developed dozens of data processing models and multiple generations of data acquisition devices to achieve automated control of data quality and comprehensively enhance the efficiency of data collection.

At the hardware level, the company has independently developed and continuously optimized robot bodies adapted for control by multimodal large models, better meeting the demands for precise manipulation and stable operation in open environments. Currently, these proprietary robot bodies have been deployed in multi-step complex task scenarios.

Notably, since late last year, Zibianliang Robotics has achieved breakthroughs in multimodal output and embodied chain-of-thought reasoning, predating Gemini Robotics’ March announcement of its chain-of-thought research findings. The company’s model architecture, centered on multimodal output, enables end-to-end information fusion, facilitating efficient alignment across visual, linguistic, and motor channels. This significantly enhances the model’s contextual reasoning and self-feedback capabilities in ultra-long-sequence tasks. The company has also made continuous progress in dynamic environment perception, real-time task planning, and long-horizon task execution. These breakthroughs and advancements further bolster robots’ autonomous decision-making capabilities and operational efficiency in complex, open-ended environments.

The company’s core team brings together experts from world-renowned artificial intelligence and robotics laboratories, as well as scholars from top-tier universities in China and abroad, dedicated to advancing innovation and development in embodied AI and robotics. Core team members include the researcher who first proposed the Attention mechanism (the cornerstone of the Transformer architecture), experts from leading international robotics laboratories, technical leads behind China’s first generation of hundred-billion-parameter large language models, and senior experts in robotics hardware.

Wang Qian, Founder and CEO, graduated from Tsinghua University and is one of the earliest scholars to introduce attention mechanisms into neural networks. During his doctoral studies, Wang participated in multiple Robotics Learning research projects at a top-tier U.S. robotics laboratory, with research interests spanning various frontier areas of robotics.

Wang Hao, Co-Founder and CTO, holds a Ph.D. in Computational Physics from Peking University. During his tenure at the International Digital Economy Academy (IDEA) in the Guangdong-Hong Kong-Macao Greater Bay Area, he served as the Head of Algorithms for the Fengshenbang Large Model Team. He led the release of China’s first open-source multimodal large model, “Taiyi,” the first batch of hundred-billion-parameter large language models, “Randeng,” and the hundred-billion-parameter large language model, “Jiang Ziya.”

Regarding the two rounds of financing, Wang Qian believes that the development of embodied AI relies on advantages in model algorithms in the short term, data advantages in the medium term, and product advantages as the core driver in the long term. Zibianliang has consistently deepened its accumulation in these three key areas, taking the lead in implementing practical applications within open service scenarios in China and exploring closed-loop service models across diverse contexts.

Ji Wei, Founding Managing Partner of Huaying CapitalStatement: CPT firmly believes in the future market potential of embodied AI. The core rationale for investing in Zibianliang Robotics lies in its advantages of “scarce team DNA + full-stack technological barriers + strategic scenario positioning.” The team’s ability to integrate technologies in the field of embodied AI is rare in China, leaving a profound impression on us: Founder Wang Qian is one of the earliest researchers of the global attention mechanism (the core module of Transformer); CTO Wang Hao led the development of “Jiangziya,” China’s first large-scale model with tens of billions of parameters; and the hardware team possesses comprehensive experience from prototyping to mass production. At the model level, the company has adopted an end-to-end unified large-model approach featuring a “perception-decision-execution” closed loop, achieving a success rate of over 90% in long-sequence complex tasks such as zipper fastening and clothing organization, with certain metrics surpassing international competitors. On the business front, the company has already made initial commercialization attempts in typical application scenarios, such as elderly care and health support. Embodied AI is expected to reach its “Aha moment” within the next three years, and we are optimistic about Zibianliang Robotics’ potential to lead and define the standards for the next generation of general-purpose robots.

Chen Yu, Partner at Yunqi CapitalStatement: Zibianliang is a team with an exceptionally high concentration of large-model expertise in the field of embodied AI. Since its founding in 2023, the team has consistently adhered to exploring the scaling laws of embodied AI and developing end-to-end manipulation foundation models. Today, this technical approach is widely recognized by both academia and industry as one of the most promising directions for breakthroughs. Yunqi Partners began tracking Zibianliang shortly after its establishment. After observing the team’s generalizable demos and, more importantly, recognizing their profound capabilities in large models, data engineering pipelines, and foundational infrastructure, Yunqi made its investment decision. The core driver of embodied AI is artificial intelligence. We look forward to Zibianliang’s continued breakthroughs in general-purpose foundation models, achieving the “ChatGPT moment” for embodied AI at an early date!

Zhou Sizheng, Project Lead at GF XindeStatement: Zibianliang Robotics possesses a dual heritage in both large language models (LLMs) and robotics, offering unique and innovative insights into the integration of these technologies, and has emerged as an industry-leading player in embodied LLMs. The team has independently developed comprehensive model infrastructure, advanced model frameworks, and cutting-edge algorithms. The WALL-A model has demonstrated exceptional performance far exceeding expectations in core dimensions such as model generalization and complex task processing capabilities. GF Xinde looks forward to collaborating with Zibianliang Robotics in the future, joining hands with upstream and downstream partners to deepen efforts in technology R&D, product iteration, and market expansion. By accelerating model optimization and hardware upgrades, we aim to facilitate the efficient deployment of technological achievements, bringing embodied intelligence into factory workshops and countless households at an early stage, thereby jointly ushering in a new chapter of the intelligent era.