X Square Robot
Last edited
Fact-checked
In review queue
Sources
18 citations
Revision
v5 · 3,674 words
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
X Square Robot (Chinese: 自变量机器人; pinyin: Zìbiànliàng Jīqìrén), also written as X²-Robot, is a Chinese robotics company headquartered in Shenzhen, China, that builds both general-purpose humanoid robots and the embodied AI foundation models that control them. Founded in December 2023 by Wang Qian (王潜), the company is known for the QUANTA X2 wheeled humanoid robot, the earlier Quanta X1 platform, and the WALL family of embodied foundation models, of which WALL-OSS was released as an open-source vision-language-action (VLA) model on GitHub and Hugging Face in September 2025.[1][2]
By mid-2026, X Square Robot had raised capital across a rapid succession of rounds, including four consecutive financings in 2026 that lifted its valuation past $2.8 billion (about RMB 20 billion), placing it among China's highest-valued embodied AI startups.[13] A roughly $100 million Series A in September 2025 was led by Alibaba Cloud, Alibaba's first investment in the embodied AI sector; a roughly 1 billion yuan Series A++ in January 2026 was led by ByteDance and the venture firm HongShan (formerly Sequoia China); a nearly $276 million Series B in April 2026 was led by the strategic investment arm of Xiaomi; and a Series C in June 2026 drew in IDG Capital.[3][4][13][14] The company describes itself as the only embodied AI firm in China to have secured lead-round backing at different stages from four of the country's leading technology companies: Alibaba, ByteDance, Meituan, and Xiaomi.[13]
History
Founding and early development (2023 to 2024)
X Square Robot was registered in Shenzhen's Nanshan district in December 2023. The legal entity is Variable Robotics Technology Co., Ltd., and the company's offices sit on the 31st floor of the Aerospace Innovation Building at 7013 Liuxian Boulevard in the Xili subdistrict, a corridor that is home to many of Shenzhen's AI and robotics startups.[2]
Founder Wang Qian named the company around the Chinese term zìbiànliàng (自变量), the mathematical phrase for an independent variable. He has said in interviews that the "x" carries a double meaning: the symbol for an unknown, and the ambition to act as "a variable that changes the world."[5]
In October 2024, the company trained an early in-house vision-language-action model called WALL-A. Demonstration videos showed a single model controlling robots through a wide range of unrelated tasks, including hanging laundry on a rack, preparing shaved ice, winding cables around pegs, and sorting parcels of arbitrary shape. The videos drew attention in Chinese tech press because the company was not using task-specific reprogramming between demonstrations.[6]
Series A and the QUANTA X2 (2025)
In August 2025, X Square Robot unveiled the QUANTA X2, a wheeled dual-arm humanoid robot with 62 total degrees of freedom. The company has said it developed the complete technology stack for the robot in under six months, including the chassis, dexterous hands, exoskeleton teleoperation suit, and supporting data-collection rigs.[7]
On September 8, 2025, X Square Robot announced approximately $100 million in Series A funding. The round was led by Alibaba Cloud with participation from Meituan, HongShan, INCE Capital, Legend Capital, and Legend Star. It was Alibaba's first reported investment in the embodied AI category, and was the company's seventh or eighth round since founding (Chinese-language reporting differs on the exact count because two waves of strategic investments were announced together).[3][6]
Alongside the Series A, the company open-sourced WALL-OSS, a 4-billion-parameter VLA model trained on real robot data combined with augmented generative video. The model was released on Hugging Face under the x-square-robot organization, with the initial public checkpoints labelled wall-oss-flow and wall-oss-fast.[1][8]
Series A++ and broader backing (January 2026)
On January 13, 2026, the company announced a Series A++ round of roughly 1 billion yuan (about $140 million), led by ByteDance and HongShan, with co-investment from the Beijing Information Industry Development Fund, the Shenzhen Innovation Investment Fund, the Nanshan Strategic Emerging Industries Fund, and several other regional vehicles. Chinese tech outlet QbitAI called it the largest embodied AI financing of the year and noted that the round made X Square Robot the only domestic embodied AI startup backed by ByteDance, Alibaba, and Meituan at the same time.[4]
In February 2026, the company posted a revised checkpoint wall-oss-flow-v0.1 to Hugging Face.[8]
Series B, WALL-B, and the push into homes (spring 2026)
On March 18, 2026, X Square Robot partnered with the Chinese classifieds and local-services platform 58.com to launch what the companies described as China's first home-cleaning robot service, beginning in Shenzhen and later extending to Beijing.[17] On April 22, 2026, the company unveiled WALL-B, a new embodied AI foundation model, and said its robots would enter everyday homes within 35 days of the announcement.[9]
On April 21, 2026, X Square Robot disclosed a nearly $276 million (about RMB 2 billion) Series B round led by the strategic investment arm of Xiaomi and by HongShan, with participation from returning backers including ByteDance, Alibaba, Meituan, and the Beijing Information Industry Development Fund. The company earmarked the capital for scaling household robot deployments and said it was targeting the first in-home trials by late May.[14][18]
Series C and a $2.8 billion valuation (mid-2026)
On June 29, 2026, X Square Robot announced that it had closed four consecutive financing rounds culminating in a Series C, which included IDG Capital, and that the rounds had lifted its valuation past $2.8 billion (about RMB 20 billion), among the highest of any Chinese embodied AI startup.[13] The company said the sequence made it the only embodied AI firm in China to have secured lead-round backing at different stages from four of the country's leading technology companies: Alibaba, ByteDance, Meituan, and Xiaomi.[13] Founder Wang Qian framed the strategy around the company's in-house model program: "Since day one, X Square Robot has focused on in-house development of foundation models, pursuing a challenging but necessary path."[13]
Founder and leadership
Wang Qian (王潜)
Wang Qian is the founder and chief executive officer of X Square Robot. He holds bachelor's and master's degrees from Tsinghua University and pursued a PhD at the University of Southern California, where his research focused on robot learning and human to robot interaction. Chinese-language profiles in BAAI's Zhiyuan publication describe him as one of the earliest researchers to introduce the attention mechanism into neural networks; a 2014 paper of his appeared at the same conference as Google's contemporaneous attention work, which later fed into the Transformer architecture that powers modern large language models.[5]
Before founding X Square Robot, Wang ran a quantitative hedge fund, where he built end-to-end statistical trading models. He has described the move from finance to robotics as a personal reorientation: quantitative trading, in his framing, is fundamentally about extracting profit from existing markets, whereas embodied AI is about building something that affects how people live. On November 2, 2025, he was selected as a torchbearer for the 15th National Games held in Shenzhen.[5]
Wang Hao (王昊)
The chief technology officer is Wang Hao, who holds a PhD in computational physics from Peking University and previously led the large language model team at a Chinese AI research institute. He oversees the WALL model program, training infrastructure, and the company's data factory. At the WALL-B launch he described the model's core idea in developmental terms: "We train vision, language, action and prediction in the same network from day one. Human infants do not learn to see, move and communicate in isolated stages. They learn by integrating perception and action at the same time."[4][9]
Products
QUANTA X1
The Quanta X1 is X Square Robot's first wheeled bimanual robot, marketed for research, education, and light commercial automation. It served as an internal data-collection platform for the WALL model family before QUANTA X2 superseded it.[2]
QUANTA X2
The QUANTA X2 is the company's wheeled dual-arm humanoid robot, unveiled in August 2025.
| Specification | Details |
|---|---|
| Height | 172 cm |
| Weight | approximately 95 kg |
| Total degrees of freedom | 62 |
| Arm degrees of freedom | 7 per arm |
| Hand degrees of freedom | 20 per hand (dexterous, five-fingered) |
| Chassis | 6-degree-of-freedom omnidirectional, wheeled |
| Top base speed | approximately 1 m/s (about 3.6 km/h) |
| Per-arm payload | approximately 6 kg |
| Reach | approximately 75 cm per arm |
| Battery runtime | approximately 2 hours per charge, task dependent |
| Sensors | 2D LiDAR, ultrasonic, RGB-D cameras |
| Onboard model | WALL-A / WALL-OSS / WALL-B family |
| Indicative price | reported around $80,000 USD |
The QUANTA X2 pairs an omnidirectional wheeled base with two 7-DoF arms ending in 20-DoF dexterous hands. The hands feature tactile sensing that the company says can detect subtle pressure changes; demonstrations have shown the robot performing 360-degree mop cleaning, picking up deformable items such as towels, and handling small objects with reported sub-millimeter repeatability for assembly-style work.[7][10]
The robot ships with a modular tool clamp that lets the hands swap between attachments such as mops or cleaning pads. Full-body teleoperation, captured through the company's exoskeleton suits, doubles as both a control method and a way to generate paired demonstration data for model training.[1]
ArtiXon Hand
In parallel with the QUANTA X2 launch, X Square Robot productized the dexterous hand used on the robot under the name ArtiXon Hand. The hand has five fingers and 20 degrees of freedom, and is offered as a standalone component for research labs and other humanoid OEMs.[2]
WALL embodied foundation models
Model lineage
The WALL series is X Square Robot's family of embodied foundation models. The architecture combines a vision-language-action model with world models, an approach the company describes as letting the model predict the consequences of its own actions before executing them.[1]
| Version | Released | Key facts |
|---|---|---|
| WALL-A | October 2024 (internal demos); production use through 2025 | First in-house VLA model; covers multi-task manipulation, including laundry, food prep, cable winding, parcel sorting |
| wall-oss-flow | September 9, 2025 | Open-source 4B-parameter checkpoint, flow-matching action head |
| wall-oss-fast | September 9, 2025 | Open-source 4B-parameter checkpoint, lower-latency variant |
| wall-oss-flow-v0.1 | February 2026 | Updated 4B-parameter checkpoint on Hugging Face |
| WALL-B | April 22, 2026 | First model on the World Unified Model architecture; targets unstructured home environments |
| WALL-OSS-0.5 | May 28, 2026 | Open-source 4B VLA (3B VLM backbone plus Mixture-of-Transformers action expert); full training code released |
| WALL-WM | 2026 | World model for event-level prediction across vision, language, and action |
The libero_all training dataset, an aggregated set used for benchmarking, has also been posted to the X Square Robot Hugging Face organization, where it has accumulated more than 270,000 downloads by mid-2026.[8]
Architecture and training
X Square Robot published a technical report for WALL-OSS, titled "Igniting VLMs toward the Embodied Space," on arXiv on September 16, 2025.[16] The paper describes an InternVL-based vision-language model backbone extended with a shared attention backbone and task-routed feed-forward networks, layered chain-of-thought reasoning for high-level planning, and a flow-matching action head for continuous control output, trained as a single cross-embodiment model that can drive multiple robot morphologies. The company says it trains in three stages: high-level reasoning, fine-grained motor control, and a final fusion stage that ties the two together.[1][16]
A stated design goal is to mitigate two long-standing problems in robot learning. The first is catastrophic forgetting, where training on new tasks erases knowledge of older ones. The second is modal decoupling, where vision, language, and action drift out of sync in long-horizon tasks. The shared backbone plus task-routed feed-forward layers are intended to keep the model's reasoning and control circuits anchored to a common representation.[16]
What is the World Unified Model (WUM)?
At the April 22, 2026 launch of WALL-B, X Square Robot introduced the World Unified Model (WUM), a training framework that optimizes vision, language, action, and physical prediction jointly within a single network from the start, rather than training separate perception, language, and control modules and stitching them together. The company says this joint optimization lets physical prediction, including force, friction, and collision dynamics, emerge inside the model itself, which it argues is better suited to unstructured homes than the repetitive settings of a factory.[9]
Wang Qian summarized the difference this way: "Robots in factories and robots in homes are fundamentally different. In factories, they repeat the same action 10,000 times. In a home, they may need to perform 10,000 different actions, each in a different context."[9] The company also described a world-model extension, WALL-WM, focused on event-level prediction that aligns language, vision, and action data.[13]
How does WALL-OSS-0.5 perform?
On May 28, 2026, X Square Robot open-sourced WALL-OSS-0.5, a 4-billion-parameter VLA model pairing a 3B vision-language model backbone with a Mixture-of-Transformers action-expert layout and a Vision-Aligned RVQ action tokenizer. The company released the full stack, including model weights, training code, training recipes, and optimizer implementations, an end-to-end reproducibility level it noted most commercial and academic VLA releases do not match.[15]
The release emphasized zero-shot performance before any task-specific fine-tuning. Across a 17-task real-robot evaluation, WALL-OSS-0.5 cleared 80% autonomous completion on four tasks without any post-training.[15]
| Real-robot task (zero-shot, no fine-tuning) | Completion score |
|---|---|
| Block sorting | 100 |
| Fruit sorting | 96 |
| Ring stacking | 86 |
| Rope tightening | 82 |
WALL-OSS-0.5 also introduced DMuon, a distributed variant of the Muon optimizer for VLA co-training that partitions matrix-level Newton-Schulz computation across sharded parameters. The company reported that DMuon reduces end-to-end optimizer overhead by up to 100 times relative to a naive implementation.[15]
Data factory and teleoperation
Wang Qian has said that the company runs what he calls a "large-scale data collection factory" that produces proprietary teleoperation data on a continuous basis. Iteration cycles are described as roughly two to three months between major model checkpoints. The supporting hardware includes wearable exoskeletons and rigs based on the Universal Manipulation Interface (UMI), a low-cost handheld gripper design originally proposed by Stanford researchers, which the company has integrated into its own pipeline.[1]
Technology and approach
X Square Robot's distinguishing characteristic is its "full-stack" position: rather than buying robot hardware from a third party or licensing a foundation model, it builds both. The relevant capability areas are summarized below.
| Capability area | What the company builds in-house |
|---|---|
| Robot platform | QUANTA X1 and X2 humanoid bases, omnidirectional chassis, body frames |
| Dexterous manipulation | ArtiXon Hand, 20-DoF five-finger hand with tactile sensing |
| Foundation models | WALL-A and WALL-B (proprietary) and WALL-OSS (open source) |
| Data infrastructure | Exoskeleton teleoperation suits, UMI-style handheld grippers, large-scale demonstration pipelines |
| World models | World Unified Model and WALL-WM predictive modules that simulate task outcomes before action execution |
| Open source | WALL-OSS and WALL-OSS-0.5 checkpoints, training code, and the libero_all dataset on Hugging Face |
The company positions WALL as an embodied counterpart to text-only large language models, arguing that embodied intelligence requires a separate model family because it must handle continuous physical interactions instead of discrete tokens. This framing is similar to claims made by Figure AI, Physical Intelligence, Skild AI, and other international competitors, but X Square Robot's approach leans harder on world-model conditioning and on producing its own data through teleoperation rather than relying on web-scraped video alone.[1][6]
Commercial deployment
The company has confirmed in interviews that it was generating revenue from commercial sales before its major funding rounds, an unusual position for an embodied AI startup of its age. CNBC reporting from September 2025 listed three deployment categories.
| Setting | Reported use case |
|---|---|
| Schools | Educational demonstrations and classroom-scale automation |
| Hotels | Guest service and routine facility management |
| Retirement homes | Light elderly-care assistance and companionship tasks |
Wang Qian told CNBC that the company planned to expand into Japan and Singapore as early international markets, citing aging populations and labor shortages as drivers of demand for service robots. Small-scale shipments of QUANTA X2 hardware were targeted for the end of 2025.[3][11]
When did X Square Robot start putting robots in homes?
In 2026 the company moved aggressively toward consumer homes. On March 18, 2026, it partnered with the local-services platform 58.com to launch what the two firms called China's first home-cleaning robot service, initially in Shenzhen and later expanded to Beijing. The service pairs a professional cleaner with an X Square robot in a robotics-as-a-service model: the human handles detailed, judgment-based work while the robot autonomously performs structured tasks such as wiping surfaces, tidying items, and collecting debris.[17]
Alongside the April 22, 2026 launch of WALL-B, the company said its robots would enter everyday homes within 35 days.[9] It followed through with an "X Family Member Program" that placed robots with volunteer households for stretches of up to one month, where the machines performed routine chores and generated operational data used to keep improving the models.[13] Chinese reporting at the time of the Series B described the company as targeting its first in-home trials by late May 2026.[18]
Funding history
| Round | Amount (approx.) | Date | Lead investor(s) | Notable co-investors |
|---|---|---|---|---|
| Angel and seed (combined) | undisclosed (multiple rounds) | 2023 to early 2025 | Legend Star, Legend Capital | Various |
| Pre-A and Pre-A+ | hundreds of millions yuan (combined) | early 2025 | Lightspeed China, Junlian Capital | Various |
| Series A (independent round) | undisclosed | mid-2025 | Meituan (exclusive lead) | Existing backers |
| Series A (combined announcement) | approximately $100 million USD | September 2025 | Alibaba Cloud | Meituan, HongShan, INCE Capital, Legend Capital, Legend Star |
| Series A+ | approximately 1 billion yuan | late 2025 | Alibaba Cloud | CNKI Investment, others |
| Series A++ | approximately 1 billion yuan (about $140 million USD) | January 2026 | ByteDance, HongShan | Beijing Information Industry Development Fund, Shenzhen Innovation Investment Fund, Nanshan Strategic Emerging Industries Fund |
| Series B | nearly $276 million USD (about RMB 2 billion) | April 2026 | Xiaomi strategic investment arm, HongShan | ByteDance, Alibaba, Meituan, Beijing Information Industry Development Fund |
| Series C | undisclosed; valuation over $2.8 billion (about RMB 20 billion) | June 2026 | IDG Capital | HongShan, Xiaomi, and returning investors |
By mid-2026, after the Xiaomi-led Series B and the Series C, Chinese tech press was describing X Square Robot as the only domestic embodied AI company to have won lead-round backing from four of China's leading technology companies: Alibaba, ByteDance, Meituan, and Xiaomi. The four consecutive 2026 rounds lifted its valuation past $2.8 billion (about RMB 20 billion), and the rapid succession of financings, roughly ten in under three years, reflects a heavy concentration of capital into a small group of frontline embodied AI startups.[4][12][13]
Reception and industry context
Industry coverage in 2025 and 2026 has placed X Square Robot inside a tight cohort of Chinese embodied AI companies that also includes Unitree Robotics, Galbot, AgiBot, and Fourier Intelligence. Within that group, the company is typically described as the most software-forward operator, on the strength of WALL-OSS and the public Hugging Face releases.[10]
The Robot Report and Yicai Global have both flagged the company's open-source positioning as a counterweight to the closed-model strategies of U.S. competitors such as Figure AI and Physical Intelligence. Open-sourcing a 4-billion-parameter VLA checkpoint gave WALL-OSS one of the larger publicly available action-conditioned models at the time of release, and the May 2026 WALL-OSS-0.5 release extended that lead by publishing full training code alongside the weights.[1][8][15] By June 2026, the four consecutive 2026 rounds and the resulting $2.8 billion-plus valuation had made X Square Robot one of the best-capitalized embodied AI startups in China.[13] Wang Qian has said he expects 2026 to be the first year in which embodied AI demonstrates clear positive return on investment in commercial settings, a claim that has not yet been independently verified.[5]
See also
- QUANTA X2
- Embodied AI
- Vision-language-action models
- World model
- Humanoid robot
- Robotics as a service
- Humanoid robot deployments
- Alibaba Cloud
- ByteDance
- Xiaomi
- Figure AI
- Physical Intelligence
- Unitree Robotics
References
- The Robot Report. "X Square Robot secures $140M in funding for AI foundation models." January 2026. https://www.therobotreport.com/x-square-robot-secures-140m-in-funding-for-ai-foundation-models/ ↩
- X Square Robot, official website. "About X Square Robot." https://x2robot.com/en (accessed 2026). ↩
- CNBC. "Alibaba leads $100 million investment in Chinese humanoid robot startup." September 8, 2025. https://www.cnbc.com/2025/09/08/alibaba-leads-100-million-investment-in-chinese-humanoid-robot-startup.html ↩
- QbitAI (量子位). "具身智能开年最大融资，字节红杉领投10亿." January 2026. https://www.qbitai.com/2026/01/369147.html ↩
- BAAI Hub (智源社区). "2026年见分晓！自变量王潜揭秘具身智能唯一破局之路丨智源专访." 2026. https://hub.baai.ac.cn/view/51046 ↩
- The Robot Report. "X Square Robot debuts foundation model for robotic butler after Series A round." September 2025. https://www.therobotreport.com/x-square-robot-debuts-foundation-model-embodied-ai-100m-series-a/ ↩
- Aparobot. "Quanta X2: Robot Details, Use Case and Specifications." 2025. https://www.aparobot.com/robots/quanta-x2 ↩
- Hugging Face. "x-square-robot organization page." 2026. https://huggingface.co/x-square-robot ↩
- PR Newswire. "X Square Robot Unveils New Embodied AI Model, Says Robots Will Arrive in Homes in 35 Days." April 22, 2026. https://www.prnewswire.com/news-releases/x-square-robot-unveils-new-embodied-ai-model-says-robots-will-arrive-in-homes-in-35-days-302751047.html ↩
- Humanoid.guide. "QUANTA X2." 2025. https://humanoid.guide/product/quanta-x2/ ↩
- Yicai Global. "China's X Square Robot Secures USD140.3 Million in Series A+ Fundraiser, Releases Open-Source Foundation Model." 2025. ↩
- Sina Finance (新浪科技). "自变量机器人，一年融4轮." February 25, 2025. https://finance.sina.com.cn/tech/roll/2025-02-25/doc-inemssqh8987214.shtml ↩
- PR Newswire. "X Square Robot Secures Four Consecutive Financing Rounds, Surpasses US$2.8 Billion Valuation in Push for Physical AI Foundation Models." June 29, 2026. https://www.prnewswire.com/apac/news-releases/x-square-robot-secures-four-consecutive-financing-rounds-surpasses-us2-8-billion-valuation-in-push-for-physical-ai-foundation-models-302813098.html ↩
- Pandaily. "Embodied AI Startup X Square Robot Raises Nearly $276M in Series B Led by Xiaomi and Sequoia China." April 2026. https://pandaily.com/embodied-ai-startup-x-square-robot-raises-nearly-276-m-in-series-b-led-by-xiaomi-and-sequoia-china ↩
- PR Newswire. "X Square Robot Open-Sources Wall-OSS-0.5, Bringing Pretrained VLA Performance Closer to Post-Training Levels." May 28, 2026. https://www.prnewswire.com/news-releases/x-square-robot-open-sources-wall-oss-0-5--bringing-pretrained-vla-performance-closer-to-post-training-levels-302784293.html ↩
- Zhai, A., et al. "WALL-OSS: Igniting VLMs toward the Embodied Space." arXiv:2509.11766. September 16, 2025. https://arxiv.org/abs/2509.11766 ↩
- PR Newswire (PR Newswire Asia). "X Square Robot and 58.com Launch China's First Home Cleaning Robot Service in Shenzhen." March 18, 2026. https://en.prnasia.com/releases/global/x-square-robot-and-58-com-launch-china-s-first-home-cleaning-robot-service-in-shenzhen-525752.shtml ↩
- Caixin Global. "X Square Robot Raises New Funds, Targets Home Trials by May." April 21, 2026. https://www.caixinglobal.com/2026-04-21/x-square-robot-raises-new-funds-targets-home-trials-by-may-102436558.html ↩
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.
4 revisions by 1 contributors · full history