Spirit AI
Last reviewed
May 11, 2026
Sources
16 citations
Review status
Source-backed
Revision
v3 · 2,500 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 11, 2026
Sources
16 citations
Review status
Source-backed
Revision
v3 · 2,500 words
Add missing citations, update stale details, or suggest a clearer explanation.
Spirit AI (officially Spirit AI (Hangzhou) Technology Co., Ltd.; Chinese: 千寻智能; Qianxun Intelligence) is a Chinese robotics company headquartered in Hangzhou, Zhejiang, China, specializing in embodied intelligence robots and vision-language-action models. Founded in early 2024 by Han Fengtao and AI scientist Gao Yang, Spirit AI develops general-purpose humanoid robots and the foundation models that control them. The company's flagship products include the Moz1 humanoid robot and the Spirit v1.5 embodied intelligence foundation model, which achieved the top ranking on the RoboChallenge real-world robotics benchmark in January 2026, surpassing the pi0.5 model developed by US-based Physical Intelligence.
By April 2026, the company had raised more than 3.3 billion yuan (roughly 460 million US dollars) across six rounds in about two years, with a post-money valuation above 10 billion yuan. Backers include Shunwei Capital (Lei Jun's fund), Yunfeng Capital (Jack Ma's fund), JD.com, Sequoia China, Prosperity7 Ventures, CATL, Huawei, Xiaomi, TCL Capital, and 360 Capital.
Spirit AI was founded in early 2024 (registration documents and most English-language sources cite February 2024; several Chinese profiles list January 2024) by Han Fengtao, who previously co-founded the collaborative-robot maker Rokae Robotics (also transliterated as LuoShi or Luoshi Robotics) in 2015 and served as its chief technology officer.[1][2] At Rokae, Han led the development of roughly 100 industrial robot models and shipped more than 20,000 units, giving Spirit AI an unusually mass-production-literate founding team for an embodied-AI startup.[2]
Han was joined by Gao Yang, a UC Berkeley PhD who trained under Pieter Abbeel and is an assistant professor at Tsinghua University's Institute for Interdisciplinary Information Sciences, where he directs the Tsinghua Vision and Embodied Intelligence Laboratory. Gao is the company's chief scientist and led earlier academic work on the ViLa and CoPa models for vision-language grounding in manipulation. A third co-founder, Zheng Lingyin, serves as chief operating officer.[2][3]
The company's first robot, the Moz0, appeared in July 2024 as an internal development platform. In March 2025, Spirit AI publicly demonstrated its Spirit v1 vision-language-action model by showing a robot completing a full clothes-folding sequence, then widely described as the first Chinese demonstration of long-horizon flexible-object manipulation by an end-to-end VLA system.[3] The commercial-grade Moz1 humanoid robot followed in June 2025.[4]
In the second half of 2025 the company expanded rapidly, growing its data-collection workforce and signing a strategic partnership with JD.com under which Moz robots were deployed at JD MALL stores to demonstrate teleoperated barista work using JD's JoyAI and JoyInside platforms. The demonstrations doubled as a data-collection loop, capturing teleoperation trajectories and force feedback to fine-tune the company's models.[5][6]
On December 17, 2025, Spirit AI's robot, branded "Xiaomo" in Chinese and "Moz" in English, was deployed on what CATL described as the world's first humanoid-robot power-battery production line at the company's Luoyang base in Henan province.[7][8] Five weeks later, in January 2026, Spirit v1.5 topped the RoboChallenge real-world robotics benchmark with a score of 66.09, surpassing the pi0.5 model from Physical Intelligence; Spirit AI simultaneously open-sourced the model and supporting resources.[9][10] In February and April 2026, Spirit AI announced two consecutive large funding rounds totaling roughly 3 billion yuan, pushing its valuation past 10 billion yuan and making it one of the highest-valued private embodied-AI companies in China.[11][12]
The Moz1 is Spirit AI's first commercial-grade humanoid robot, unveiled in June 2025. It is positioned for industrial, logistics, and service-sector tasks rather than as a household consumer product.
| Specification | Details |
|---|---|
| Form factor | Bipedal humanoid |
| Degrees of freedom | 26 (excluding dexterous end effectors) |
| Joints | Self-developed integrated force-controlled actuators |
| Power density | Highest among shipping humanoids (company claim) |
| Load-to-weight ratio | 1:1 |
| Whole-body control | Self-developed high-precision, high-speed WBC algorithm |
| Teleoperation | Zero-latency full-body, paired with proprietary capture rig |
| Data device | Multi-dimensional collection unit for hourly model iteration |
| Target applications | Logistics, battery manufacturing, retail service, factory testing |
| Released | June 2025 |
The Moz1's joints combine motor, gearbox, encoder, and torque sensor into a single module the company designs in-house. Han Fengtao has argued that a robotics company that does not build its own VLA model cannot tell what "good hardware" looks like; the actuator specification reflects that view, with joints tuned for the contact-rich tasks the Spirit foundation models try to learn.[2][4]
Moz1 ships with a wearable data-capture system that Spirit AI uses to collect demonstrations across factories, homes, and labs. The fifth-generation rig, in use by early 2026, captured what the company described as 95 percent usable data at roughly one-tenth the cost of conventional teleoperation. As of February 2026 Spirit AI had accumulated more than 200,000 hours of real-world interaction data from internet video, teleoperation, and wearables, with a stated 2026 target above 1 million hours and a collection team scaling to about 1,000 people.[11][13]
Spirit v1, released in March 2025, was a vision-language-action model demonstrated on long-horizon manipulation tasks such as folding garments. Spirit v1.5, open-sourced in January 2026, is the successor and the model that took the top RoboChallenge ranking.
| Metric | Spirit v1.5 |
|---|---|
| Architecture | Unified end-to-end vision-language-action |
| Parameter scale | Under 10 billion |
| RoboChallenge total score | 66.09 |
| Task success rate | 50.33% (only model above 50% at release) |
| Ranking | First on RoboChallenge real-world leaderboard (January 2026) |
| Closest competitor | pi0.5 (Physical Intelligence) |
| License | Open source (model and supporting resources) |
| Training signal | Pretraining on web video, fine-tuning on factory teleoperation, reinforcement learning |
In explaining the architecture, Han Fengtao said Spirit v1.5 "integrates perception, reasoning and action into an end-to-end system, reducing errors associated with established modular approaches."[9][10] The team has publicly noted a scaling-law pattern: roughly a tenfold increase in training data yields a step-function improvement of about nine points on internal benchmarks, which is the rationale for the aggressive million-hour data target.[6][13]
Spirit AI's stack is full-stack by design, mixing foundation-model research with motor and joint engineering. Core areas:
Earlier academic work by Gao Yang, including the ViLa visual-language model and the CoPa component-constraint model, feeds into Spirit AI's perception stack.[3]
The largest production deployment is at CATL's Zhongzhou base in Luoyang, Henan, announced on December 17, 2025 and described by CATL as the world's first humanoid-robot power-battery production line.[7][8][14] Moz robots handle End-of-Line (EOL) and Direct Current Resistance (DCR) tests, which involve plugging high-voltage test connectors into specific points on a battery pack. The job is hazardous when performed by humans because the test plugs carry hundreds of volts.
Reported performance from CATL and Spirit AI:
| Metric | Reported value |
|---|---|
| Plug-in success rate | Consistently above 99% |
| Cycle time | Comparable to skilled human workers |
| Daily workload | Roughly triple a single human worker, due to continuous operation |
| Battery models handled | Multiple, with on-the-fly position adjustment |
| Secondary role | Patrol inspection during idle intervals |
The deployment ran on CATL-supplied batteries and used the Spirit VLA model to adjust grip and pose in real time when connector positions varied between battery packs.[7][8][14]
Under the 2025 strategic partnership with JD.com, Moz robots were placed in JD MALL physical stores, where they performed barista tasks including cup pickup, bean grinding, espresso extraction, and latte service. The robots used JD's JoyAI large model and JoyInside teleoperation platform, and every interaction fed multimodal sensor data, joint trajectories, and force feedback back into Spirit AI's training pipeline. Han Fengtao has framed the deployment as much as a data flywheel as a product launch.[5][6]
Spirit AI's robots participated in the 2025 Intelligent Robot Competition in China, completing tasks such as desktop tidying, waste disposal, and microwave operation. In January 2026, Spirit v1.5 topped the RoboChallenge leaderboard, evaluated across 30 real-world manipulation tasks including object placement, target recognition, and tool use; it was the only model to clear the 50 percent task-success-rate threshold at release.[2][9][10]
Spirit AI raised six rounds of equity capital between August 2024 and April 2026, totaling roughly 3.3 billion yuan (about 460 million US dollars at then-prevailing rates) and pushing the company's post-money valuation above 10 billion yuan.
| Round | Amount | Date | Lead investor(s) | Notable participants |
|---|---|---|---|---|
| Angel | ~200 million yuan | August 2024 | Shunwei Capital | Borui Capital, Fortune Capital |
| Angel+ | Undisclosed | November 2024 | Undisclosed | Existing investors |
| Pre-A | 528 million yuan | March 2025 | Prosperity7 Ventures | China Merchants Venture, GF Xinde, Jingya Capital |
| Pre-A+ | ~600 million yuan | July 2025 | JD.com | China Internet Investment Fund, Zhejiang STI Mother Fund, Huatai Zijin, Fosun Rui Zheng |
| Series B (split tranche 1) | ~2 billion yuan | February 2026 | Yunfeng Capital, Sequoia China, TCL Capital, 360 Capital | Chaos Investment, Synstellation Capital, regional government funds |
| Series B (split tranche 2) | ~1 billion yuan | April 2026 | Shunwei Capital, Yunfeng Capital | Fortune Capital, Galaxy Yuanhui, Turing Fund, Xinding Capital, Gengxin Capital, JD.com |
The February 2026 round was reported in English-language press as roughly 290 million US dollars and characterized by PR Newswire and other outlets as a fundraise to scale Spirit AI's "dirty data" pipeline.[11][13][15] The April 2026 round, co-led by Lei Jun's Shunwei Capital and Jack Ma's Yunfeng Capital, was widely reported as the first time the two funds jointly led an embodied-AI investment.[12][13]
Strategic investors with industry rather than pure-financial motivations include CATL (battery production deployments), JD.com (retail and logistics), Huawei and Xiaomi (consumer-electronics and silicon ecosystem), TCL (appliance manufacturing), and Inovance Technology (industrial automation). Founder Han Fengtao reportedly held about 24.86 percent of the company directly and 16.39 percent indirectly as of mid-2025.[2][3]
| Name | Role | Background |
|---|---|---|
| Han Fengtao | Founder and CEO | Former co-founder and CTO of Rokae Robotics (Luoshi Robotics); led development of roughly 100 industrial robot models and shipped over 20,000 units |
| Gao Yang | Co-founder and chief scientist | UC Berkeley PhD under Pieter Abbeel; assistant professor at Tsinghua University and director of the Tsinghua Vision and Embodied Intelligence Laboratory; lead author on ViLa and CoPa |
| Zheng Lingyin | Co-founder and COO | Industrial-robotics commercialization veteran with overseas-expansion experience |
Core team members are drawn from UC Berkeley, Carnegie Mellon University, Tsinghua University, Peking University, Xiaomi, Huawei, and Tencent.[3]
Spirit AI splits its resources heavily in favor of model and data work, with Han saying the company puts roughly 80 percent of resources into the data system and model rather than hardware program-management.[6][13] The stated commercial path is to enter narrow vertical markets first (battery assembly, retail service, laundry-style flexible manipulation) and use them as data flywheels for a more general model, with broader deployment of service robots expected within two to three years of Spirit v1.5.[9][10]
Han has compared the state of embodied AI in 2026 to LLMs in 2023: a moment when raw funding and benchmark performance become threshold requirements rather than differentiators.[2] The 2026 commercial targets reported by executives included 100 to 150 million yuan in revenue and roughly 200 to several-hundred Moz robot shipments, with larger volumes contingent on customer-side production-line redesigns.[2][16]
This article covers the Chinese robotics company Spirit AI / Qianxun Intelligence. It is unrelated to Spirit AI Ltd., a UK-based interactive-storytelling and dialogue-tools company known for the Character Engine middleware used in some games. The two companies share an English brand name but have no corporate, technical, or personnel overlap.