OpenLoong
Last reviewed
May 11, 2026
Sources
15 citations
Review status
Source-backed
Revision
v5 · 2,345 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 11, 2026
Sources
15 citations
Review status
Source-backed
Revision
v5 · 2,345 words
Add missing citations, update stale details, or suggest a clearer explanation.
| OpenLoong | |
|---|---|
| General information | |
| Full name | National and Local Co-built Humanoid Robotics Innovation Center / OpenLoong Open-Source Community |
| Operated by | Humanoid Robotics (Shanghai) Co., Ltd.; Shanghai Humanoid Robotics Manufacturing Innovation Center; OpenAtom Foundation |
| Established | December 2023 (company); May 2024 (national center designation); June 2024 (open-source community launch) |
| Headquarters | Zhangjiang Robot Valley, Shanghai, China |
| Industry | Robotics, Humanoid robots, Open-source AI |
| Key people | Xu Bin (general manager); Jiang Lei (chief scientist) |
| Registered capital | 1 billion yuan (about USD 140 million) |
| Products | Qinglong humanoid robot series; Gewu simulation platform; MindLoongGPT |
| License | Apache License 2.0 (most repositories) |
| GitHub | github.com/loongOpen |
| Website | openloong.org.cn |
OpenLoong is a Chinese government-backed open-source humanoid robot initiative operated by the National and Local Co-built Humanoid Robotics Innovation Center in Shanghai. The center was jointly established in December 2023 by state-owned enterprises and industry-leading companies, with a registered capital of 1 billion yuan (about USD 140 million). In May 2024, China's Ministry of Industry and Information Technology and the Shanghai Municipal Government officially designated the center as the country's "National and Local Co-built Humanoid Robotics Innovation Center," giving it national-level status in coordinating humanoid robot research and development across China.[1][2][3]
The OpenLoong community provides open-source hardware designs, software packages, datasets, and development tools for the Qinglong (青龙, "green dragon") series of full-sized humanoid robots, with the stated goal of accelerating innovation and collaboration in the global robotics and artificial intelligence communities. The project is stewarded by the OpenAtom Foundation, which also hosts other major Chinese open-source projects such as OpenHarmony and openEuler.[4][5]
In December 2023, Humanoid Robotics (Shanghai) Co., Ltd. and the Shanghai Humanoid Robotics Manufacturing Innovation Center were officially established as a new research and development institution jointly founded by state-owned backbone enterprises and industry-leading firms. The center is located in the Zhangjiang Robot Valley in Shanghai's Pudong New Area and focuses on core technology development, product engineering, and application ecosystem construction for full-sized humanoid robots. Xu Bin serves as general manager and Jiang Lei serves as chief scientist.[1][2]
In May 2024, China's Ministry of Industry and Information Technology and the Shanghai Municipal Government formally designated the Shanghai facility as the country's "National and Local Co-built Humanoid Robotics Innovation Center," elevating it to a national-level institution. This designation gave the center a coordinating role across China's six regional humanoid robot innovation hubs, with the explicit mission of building an open-source ecosystem to complement work happening in Beijing, Shenzhen, Hangzhou, Chengdu, and elsewhere.[2][3]
In June 2024 the center launched the OpenLoong open-source community website, described in Chinese state media as the world's first open-source community platform devoted to full-sized general-purpose humanoid robots.[6] One month later, at the World Artificial Intelligence Conference (WAIC) 2024 in Shanghai (held 4 to 7 July 2024), the center unveiled the Qinglong humanoid robot, presented as China's first full-sized, open-source, general-purpose humanoid robot. The name "Qinglong" was chosen because 2024 is the Year of the Dragon in the Chinese zodiac, and chief scientist Jiang Lei announced that the center would introduce a new humanoid robot model annually, each named after one of the twelve Chinese zodiac animals.[7][8]
At WAIC 2024 the center released the Qinglong V1.0 hardware structure and design parameters publicly through the OpenLoong website, with associated software packages and embodied intelligence stacks following over the next year. General manager Xu Bin stated that the full stack of humanoid robot technology, including the "brain," "cerebellum," and limb control, would be released to the public as open source.[8]
In January 2025 the innovation center opened a heterogeneous humanoid robot training base in the Zhangjiang area of Shanghai, occupying over 5,000 square meters and capable of housing up to 100 humanoid robots from different manufacturers at the same time. Robots from companies including Fourier Intelligence and AgiBot signed up to train in more than ten typical application scenarios such as industrial assembly, household services, and logistics. Jiang Lei described the facility as "a factory that produces data," with sensor logs from all participating robots collected, annotated, and fed back into the OpenLoong open-source datasets. The Chinese government has projected that at least 1,000 humanoid robots will be operating in state-level training grounds of this type by 2027.[6][9][10]
In February 2025 the center hosted a forum on open-source humanoid robotics, where Xu Bin announced that 2025 plans included expanding the community capability stack and tightening cooperation with the OpenAtom Foundation. Industry standards work also accelerated during this period; the center led drafting on more than 80 industry standards for the sector, including the country's first batch of embodied intelligence standards released in Shanghai.[6][11]
In March 2025 the center partnered with Shanghai University and Tsinghua University to release the Gewu (格物) embodied intelligence simulation platform, fully open-sourced through the OpenLoong community. Gewu provides "one-click training" and "one-click migration" workflows for developers, and its general reinforcement learning framework with automated model adaptation allows algorithms to be trained across more than 100 mainstream commercial robots without rewriting code for each platform.[5]
On 5 August 2025 the center released the third generation of the Qinglong robot, Qinglong V3.0, with 60 degrees of freedom compared with 43 on the original Qinglong V1.0. The V3.0 model adds a modular design that lets users swap walking, working, and perception modules depending on the task, and it integrates with the Gewu Zhizhi ("investigation of things, knowledge of nature") cooperative training platform for embodied AI models.[12]
The center has established a humanoid robot industry fund with a first-phase target of 2 billion yuan (about USD 276 million) and a long-term target of 10 billion yuan (about USD 1.4 billion). A separate community-incentive fund of roughly 3 million yuan provides per-project grants of 300,000 to 500,000 yuan to encourage external developers to build on the OpenLoong stack.[2][6]
OpenLoong's flagship products are the Qinglong full-sized humanoid robots. Public technical disclosures cover two main generations:
| Specification | Qinglong V1.0 (July 2024) | Qinglong V3.0 (August 2025) |
|---|---|---|
| Height | ~185 cm | 185 cm |
| Weight | ~80 kg | ~85 kg |
| Active degrees of freedom | 43 | 60 |
| Maximum joint peak torque | 400 N·m | 400 N·m |
| On-board computing | 400 TOPS | 400+ TOPS |
| Hand design | Dexterous, multi-finger | 5 fingers per hand, modular grippers |
| Mobility | Bipedal walking, multi-modal | Bipedal walking, multi-modal, modular base |
| Reference release | OpenLoong open-source community, July 2024 | OpenLoong + Gewu Zhizhi, August 2025 |
Qinglong is positioned as a research and reference platform rather than a commercial product. The robot supports multi-modal locomotion, perception, and manipulation, with demonstrations at WAIC 2024 showing rapid walking, dynamic obstacle avoidance, stable movement on slopes, and precision pick-and-place tasks with the dexterous hand.[7][8][13]
OpenLoong's embodied intelligence stack is organized into modules with distinctive Chinese mythological names:
| Module | Role | Notes |
|---|---|---|
| Suzaku (朱雀) | "Brain" / cognitive layer | High-level task understanding and planning |
| Xuanwu (玄武) | "Cerebellum" / motor control | Whole-body coordination and balance |
| MindLoongGPT | Multi-modal language model | Developed with Fudan University; generates motion from text, voice, or video |
| Gewu | Simulation and training platform | Cross-platform RL training across 100+ commercial robots |
MindLoongGPT, a roughly 7-billion parameter multi-modal model developed jointly with the School of Information Science and Technology at Fudan University, lets users issue natural-language commands such as "wave" or upload a reference video, and the model parses the semantics and generates coherent motion trajectories for the Qinglong robot. The Suzaku brain and Xuanwu cerebellum modules work together with MindLoongGPT to translate language and perception into joint-level commands.[2][13]
OpenLoong's code is published on GitHub under the loongOpen organization, with most repositories licensed under Apache 2.0. As of 2025 the organization hosts more than 30 repositories; the most active include:
| Repository | Description | Language |
|---|---|---|
| OpenLoong | Top-level project hub covering cloud OS, embodied cerebellum, hardware, and data | Mixed |
| OpenLoong-Dyn-Control | Whole-body dynamics control framework based on MPC and WBC, running in MuJoCo | C++ |
| OpenLoong-Hardware | Hardware system, schematics, and parameters for the Qinglong robot | Mixed |
| Openloong-brain | Large-scale skill scheduling framework for humanoid robots | Mixed |
| Unity-RL-Playground | Reinforcement learning and embodied intelligence sandbox in Unity | C# |
| loong_driver_sdk | Low-level driver SDK for OpenLoong motors and sensors | C++ |
| LoongMarathonNav | Navigation stack used in the Beijing humanoid marathon entry | C++ |
| MiniLoong | Smaller-scale humanoid platform for education and research | C |
OpenLoong-Dyn-Control is the most-starred repository; it ships with three demonstration controllers (walking, jumping, and blind obstacle stepping) and uses the MuJoCo simulator, the Pinocchio dynamics library, and Eigen for linear algebra. The control framework combines Model Predictive Control for trajectory optimization with Whole-Body Control for instantaneous joint commands, and the codebase is self-contained so that users can clone, build, and run the demos without installing external dependencies. Some active development has migrated to the AtomGit platform run by the OpenAtom Foundation at atomgit.com/openloong, with GitHub kept as a public mirror.[4][14]
Released in March 2025, the Gewu platform addresses one of the harder problems in humanoid robotics: training a policy that has to work on many physically different robot bodies. Its automated model adaptation pipeline reads the URDF or equivalent kinematic description of a target robot, then retargets a trained policy onto that robot's joint layout without manual rewriting. The platform supports more than 100 commercial humanoid platforms from Chinese and international vendors, and the same one-click migration mechanism is used to bridge between Gewu's simulation and the physical Qinglong robot deployed at the Zhangjiang training base.[5]
OpenLoong serves as a shared research and development infrastructure for China's humanoid robotics industry rather than a commercial robotics manufacturer. The center operates as an R&D and integration hub, providing shared laboratories, testbeds, and demonstration spaces where universities, startups, and manufacturers can co-develop perception, control, and actuation technologies. This model is closer to a national laboratory than to a typical robotics company, and the Apache 2.0 licensing of most code makes it usable by commercial firms without copyleft obligations.[2][6]
China has established six major humanoid robot innovation centers nationwide, with the Shanghai center (OpenLoong) being the national-level entity focused on open-source ecosystem development. Each of the other regional centers has a different emphasis: Beijing's hub centers on open platforms and the Beijing Humanoid Robot Innovation Center (X-Humanoid), Hangzhou on consumer-facing form factors, Chengdu on lightweight designs, and so on. OpenLoong sits above and across these regional hubs in the sense that participating firms anywhere in the country can train robots at the Shanghai facility and pull from the same open-source datasets.[2]
The OpenLoong stack is often discussed alongside other open humanoid initiatives, including the Berkeley Humanoid Lite project in the United States and Open-X-Humanoid's Pelican-VL embodied brain models from the Beijing center. Compared with closed commercial platforms such as Tesla Optimus or Figure AI, OpenLoong's value proposition is that researchers can inspect and modify every layer of the stack, from CAD files and joint torque curves up through the Suzaku brain and the MindLoongGPT motion model.[4][13]
Chinese state media (China Daily, Xinhua, People's Daily) has covered OpenLoong heavily since the Qinglong unveiling at WAIC 2024, and the project is frequently cited in policy discussions of China's humanoid robotics push. International technology press, including TechNode, Interesting Engineering, AI Business, and Yicai Global, covered the launch and have continued to track the V3.0 release and the Zhangjiang training base.[7][8][9][10] Shanghai's overall humanoid robot output passed 1,000 units in 2024 according to Zhang Hongtao, deputy director of Shanghai's economy commission, with OpenLoong serving as the shared software and data backbone for many of those deployments.[6]
Independent academic adoption is visible through the Loong open platform paper published in 2025 in the journal Biomimetics, which discusses the OpenLoong system as a reference platform for practical full-sized humanoid robot research. The OpenLoong-Dyn-Control repository has been forked by independent researchers and by integration projects such as the dora-rs humanoid examples, indicating that the codebase is being used outside the original Shanghai consortium.[14][15]