Galaxea Dynamics
Last reviewed
Sources
24 citations
Review status
Source-backed
Revision
v4 · 4,074 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
24 citations
Review status
Source-backed
Revision
v4 · 4,074 words
Add missing citations, update stale details, or suggest a clearer explanation.
Galaxea Dynamics, also known as Galaxea AI (Chinese: 星海图, Xinghaitu), is a Chinese embodied AI and robotics company that builds wheeled dual-arm humanoid robots, standalone robotic arms, and the G0 family of robot-control foundation models. Founded in September 2023 by four Tsinghua University alumni, Gao Jiyang, Xu Huazhe, Hang Zhao, and Tianwei Li, and headquartered in Beijing, Galaxea raised approximately $291 million in an April 2026 Series B+ round at a valuation above CNY 20 billion (about $2.9 billion), making it one of the most highly valued embodied-AI startups in China.[4][6]
Galaxea's core products are the R1 series of wheeled, bimanual mobile manipulators (R1, R1 Pro, and R1 Lite), the A-series robotic arms, and the G0 dual-system Vision-Language-Action (VLA) model, which is trained on the Galaxea Open-World Dataset of roughly 500 hours and 100,000 demonstration trajectories collected on real robots.[8] As of April 2026 the company reported cumulative funding above $500 million and more than 40 clients including Huawei Cloud, Volkswagen, Haier, Samsung, ByteDance, Physical Intelligence, Stanford University, and MIT.[4][7] Its stated mission is "embodied intelligence at a global scale: 10 billion robots for 10 billion people."[17]
Galaxea AI was founded in September 2023 by four individuals with complementary backgrounds in autonomous driving, computer vision, and robotics research.[1][12] CEO Gao Jiyang had previously worked at Waymo and Momenta, where he led production systems for autonomous driving. Before founding the company, Gao identified a fundamental issue with the Robotaxi business model: the misalignment between high failure costs and AI's inherent need to learn through iteration. He concluded that successful AI products require "low failure costs and low data acquisition costs," which led him to pivot toward embodied intelligence and robotics.[1]
All four co-founders are Tsinghua University alumni, and several had prior working relationships. Gao and Hang Zhao collaborated at Waymo on the VectorNet paper, while Gao and Tianwei Li worked together at Momenta.[1][13] The founding team combined expertise in perception systems, robot learning, hardware engineering, and production system delivery.
In its first year, Galaxea AI focused on developing its initial hardware platform and securing early funding. The company completed multiple early-stage financing rounds backed by Baidu Ventures, GSR Ventures, and IDG Capital.[1] It released the original Galaxea R1 wheeled dual-arm humanoid robot in 2024 and began shipping units to early clients in the research and technology sectors.
In January 2025, Galaxea expanded the R1 product line with two new variants: the Galaxea R1 Pro, featuring upgraded 7-DOF arms and integrated force sensors for industrial applications, and the Galaxea R1 Lite, a compact data collection platform optimized for AI training.[11] The company also launched standalone robotic arms (the A1 and A1XY) for desktop research use.
During 2025, Galaxea closed its A4 and A5 financing rounds, raising over $100 million combined at a $700 million valuation.[2][23] Investors in these rounds included Capital Today, Meituan's Long-Z Investments, Ant Group, IDG Capital, GL Ventures, and FunPlus. In the second half of 2025, the R1 platform entered large-scale deployment, with the company targeting shipment of up to 1,000 units across domestic and international markets.[2] Revenue in 2025 was projected at tens of millions of yuan.
In late August 2025, the Galaxea team published the G0 dual-system VLA model and the Galaxea Open-World Dataset on arXiv, open-sourcing the model weights and training framework through its GitHub organization.[8][9]
In January 2026, Galaxea completed its corporate registration change, converting from a limited liability company to a joint stock company (foreign investment, unlisted). The entity was renamed "Galaxea (Beijing) Artificial Intelligence Technology Co., Ltd.," making it the first embodied AI company in China to complete this restructuring in 2026.[7] This move positioned the company for a potential future initial public offering.
The company raised approximately $144 million (CNY 1 billion) in its Series B round announced February 12, 2026, valuing the firm at roughly $1.4 billion (CNY 10 billion).[3] New investors in the round included Jinding Capital, BAIC Group Industrial Investment, and Hone Capital, while existing backers such as Cathay Capital, Meituan Dragonball, and Capital Today increased their stakes; the deal brought cumulative funding to nearly CNY 3 billion.[3] Just under two months later, on April 2, 2026, a Series B+ round brought in approximately $291 million (CNY 2 billion) from nearly 20 institutional investors, pushing the valuation above $2.9 billion (CNY 20 billion).[4][6] Cumulative fundraising surpassed $500 million, placing Galaxea among the most well-funded humanoid robotics startups in China.[5]
In March 2026, the R1 Pro was deployed at Beijing's first smart elderly care station in E-Town, demonstrating real-time clothes folding and other service tasks.[7] The company announced plans to begin volume production at the 10,000-unit level as supply chain and scenario capabilities mature through 2026.[19]
Galaxea Dynamics was co-founded by four individuals, each bringing distinct technical and operational expertise.
| Name | Role | Background |
|---|---|---|
| Gao Jiyang (高继扬) | Founder and CEO | Born 1992. Admitted to Tsinghua University through a physics competition; B.S. in Electronic Engineering from Tsinghua; Ph.D. in computer vision from the University of Southern California (USC) under Professor Ram Nevatia, completed in three years (graduated 2018). Interned at Google and SenseTime. Joined Waymo's perception team after graduation, where he co-authored the VectorNet paper. Returned to China and joined Momenta, where he improved parking success rates from 60% to 95% and delivered Navigation on Autopilot (NOA) systems to SAIC Motor.[1][12] |
| Xu Huazhe | Co-Founder and Co-Chief Science Officer | B.S. from Tsinghua University; Ph.D. from UC Berkeley's BAIR Lab under Professor Trevor Darrell; postdoctoral researcher at Stanford Vision and Learning Lab under Professor Jiajun Wu. Tenure-track Assistant Professor at Tsinghua University's Institute for Interdisciplinary Information Sciences (IIIS), where he leads the Tsinghua Embodied AI Lab (TEA Lab). Winner of best system paper at CoRL 2023.[14] |
| Hang Zhao | Co-Founder and Chief Scientist | Ph.D. from MIT (2019) under Professor Antonio Torralba. Research Scientist at Waymo (2019 to 2020), where he co-developed VectorNet with Gao Jiyang. Assistant Professor at Tsinghua University IIIS, where he leads the MARS Lab. Research focuses on multimodal machine learning, autonomous driving, and robot learning. Named to MIT Technology Review's Innovators Under 35 list.[13] |
| Tianwei Li | Co-Founder | Master's degree from University College London (UCL). Joined Momenta through campus recruitment; promoted four times in five years to Senior Director, eventually leading the SLAM team. Leads hardware and systems at Galaxea.[1] |
The company's CFO is Luo Tianqi. As of mid-2025, Galaxea had approximately 80 to 120 employees, with plans to expand to 200 by year-end.[1]
Galaxea AI has raised substantial capital in a short period, reflecting strong investor interest in the Chinese humanoid robotics sector. By April 2026 the company had completed a Series B+ round at a valuation above CNY 20 billion (about $2.9 billion) with cumulative funding exceeding $500 million.[4][5][6]
| Round | Date | Amount | Valuation | Key investors |
|---|---|---|---|---|
| Early rounds (A-series) | 2024 | Not publicly disclosed | Not disclosed | Baidu Ventures, GSR Ventures, IDG Capital |
| A4 and A5 rounds | Mid-2025 | Over $100 million combined | ~$700 million | Capital Today, Meituan Long-Z Investments, Ant Group, IDG Capital, GL Ventures, FunPlus |
| Series B | February 2026 | ~$144 million (CNY 1 billion) | ~$1.4 billion (CNY 10 billion) | Jinding Capital, BAIC Group Industrial Investment, Hone Capital, Cathay Capital, Meituan Dragonball, Capital Today |
| Series B+ | April 2026 | ~$291 million (CNY 2 billion) | ~$2.9 billion (CNY 20 billion) | Nearly 20 investors including Lens Technology, Walden International, Sixcore Investment, Times Bonile, Aero Engine Fund, CICC Capital, Puhua Capital, Hongtai Fund, GF Qianhe, Financial Street Capital, Jinpu Investment, Beijing Sci-Tech, Guoyuan Equity, Xiuyuan Capital, Hongzhang Investment, Yuhai Capital |
The investor base spans industrial capital, long-term patient funds, state-backed entities, and private equity.[4][7] Lens Technology Co., Ltd. (Shenzhen: 300433), a major consumer electronics manufacturer, participates both as an investor and as a hardware and mass production partner.[4] The Chinese government designated humanoid robotics as a national priority and established a $138 billion government fund in March 2025 for AI and robotics acceleration, providing favorable conditions for fundraising in the sector.[19]
The R1 series is Galaxea's flagship product line, consisting of three wheeled dual-arm humanoid robots that share a common design philosophy: a human-like upper body with dual manipulator arms mounted on a wheeled mobile base. This approach prioritizes practical deployability in structured indoor environments over the bipedal locomotion used by competitors like Tesla Optimus or UBTECH Robotics' Walker.
| Model | Total DOF | Arm configuration | Standing height | Max operating height | Weight | Battery | Computing | Approx. price |
|---|---|---|---|---|---|---|---|---|
| Galaxea R1 | 24 | Dual 6-DOF (A1 arms) | 170 cm | 200 cm | 96 kg | 1,680 Wh | NVIDIA Jetson AGX Orin 32GB | ~$28,000 USD |
| Galaxea R1 Pro | 26 | Dual 7-DOF (A2 arms) | 170 cm | 200 cm | 96 kg | 1,680 Wh | NVIDIA Jetson AGX Orin 32GB | ~$70,000 USD |
| Galaxea R1 Lite | 23 | Dual 6-DOF (A1X arms) | 128 cm | 170 cm | 55 kg | 720 Wh | Intel Core i9-12900HK | ~$40,000 USD |
All R1 variants feature a three-wheel omnidirectional vector steering chassis capable of Ackermann steering, translation, and spinning, with a maximum speed of 5.4 km/h.[16] The robots support robot teleoperation via isomorphic control platforms and (for the standard R1 and R1 Pro) VR headsets. The R1 and R1 Pro include 360-degree LiDAR, multiple HD cameras, speech interaction, and compatibility with ROS and ROS 2.[16]
The standard R1, released in 2024, serves as the general-purpose mobile manipulation platform. The R1 Pro, announced in January 2025, upgrades to 7-DOF A2 arms with integrated force sensors for industrial collaboration and precision tasks. The R1 Lite, also announced in January 2025, is a compact data collection platform designed for fleet deployment in research environments.[11][15]
Galaxea manufactures standalone robotic arms that are also integrated into the R1 series:
| Arm model | DOF | Key features | Used in |
|---|---|---|---|
| Galaxea A1 | 6 | Lightweight force-controlled arm with 715 mm working radius; planetary gear motors with high precision and torque; high payload-to-weight ratio | R1 (Standard) |
| Galaxea A2 | 7 | Enhanced dexterity arm with 620 mm working radius; additional degree of freedom for kinematic redundancy; integrated force sensing | R1 Pro |
| Galaxea A1X | 6 | Compact variant of A1 optimized for smaller platforms; 600 mm extended length; 4.2 kg per arm | R1 Lite |
| Galaxea A1XY | 6 | Dual-configuration desktop arm with two interchangeable forms (A1X for wrist-like agility, A1Y for lightweight low-inertia performance); designed for development, data acquisition, education, and competitions | Standalone desktop platform |
All arms are sold separately for research and development use, in addition to being integrated into the R1 platforms.
The Galaxea G1 parallel gripper is included with all R1 models. It provides 100 N of rated gripping force with a 0 to 100 mm stroke range.[16]
Galaxea has also developed the DEXO dexterous hand, which was showcased at CES 2026.[18][24] The DEXO achieves near-human dexterity with 0.1 N high-precision tactile sensing and 1 kg fingertip force per finger. The R1 Pro's arm endpoints can be swapped from grippers to dexterous hands for applications requiring anthropomorphic finger manipulation.
The B1 omnidirectional vector control chassis is offered as a standalone mobility platform for integration with third-party hardware. It uses three self-developed steering wheel modules providing 6 degrees of freedom of motion, supporting translation, spinning, and Ackermann-style steering.
The R1-T Isomorphic Remote Operation Platform is a scaled-down physical replica of the R1 that enables operators to control the robot through direct body movement mapping. The system achieves millimeter-level positional precision with millisecond-level response times and provides force feedback from the robot to the operator. It serves both as a direct control mechanism and as a data collection tool for imitation learning.
Galaxea has developed a proprietary family of AI models called G0 that serve as the intelligence layer for its robot platforms. The G0 system uses a dual-system Vision-Language-Action (VLA) architecture inspired by cognitive science, described in the September 2025 arXiv paper as "a dual-system framework that couples a Vision-Language Model (VLM) for multimodal planning with a Vision-Language-Action (VLA) model for fine-grained execution":[8]
The two systems run asynchronously at different frequencies, with System 2 operating at a slower planning cadence and System 1 executing actions at higher frequency for responsive control.[8]
G0 is trained on the Galaxea Open-World Dataset, which the paper reports comprises roughly 500 hours of high-fidelity data and about 100,000 demonstration trajectories collected entirely on the R1 Lite, a 23-DOF mobile bimanual robot.[8] The dataset covers 150 distinct task categories and 58 operational skills across 50 unique scenes and more than 1,600 unique objects, spanning four environment types: residential, retail, catering, and office.[8] Training follows a three-stage curriculum:
The authors report that "the single-embodiment pre-training stage, together with the Galaxea Open-World Dataset, plays a critical role in achieving strong performance."[8] The G0 paper was published on arXiv on August 30, 2025, with named authors including Tao Jiang, Tianyuan Yuan, Jiyang Gao, Huazhe Xu, and Hang Zhao, and the model weights were open-sourced through the GalaxeaVLA repository on GitHub (under the OpenGalaxea organization).[8][9] Model weights are also hosted on Hugging Face.
| Model | Release | Parameters | Backbone | Description |
|---|---|---|---|---|
| G0-VLA | September 2025 | Not disclosed | - | Original pre-trained dual-system VLA model |
| G0Plus | January 2026 | 3 billion | Not specified | Pre-trained VLA for multi-task manipulation; trained on 2,000+ hours of real-world robot data; includes a "Pick Up Anything" zero-shot grasping demo |
| G0Tiny | February 2026 | 250 million | SmolVLM2 | Lightweight model trained on ~1,000 hours of R1 Pro VR teleoperation data, for edge deployment on NVIDIA Jetson Orin (TensorRT inference up to 10 Hz) |
| G0.5 | June 2026 | 2 billion | Qwen3.5 2B | Pre-trained on demonstrations from 14 embodiments plus large-scale web and embodied VQA data |
G0Plus ships as a deployment-ready checkpoint that enables out-of-the-box VLA capabilities, and Galaxea's robots come with the VLA model pre-installed.[9] The June 2026 G0.5 release uses "a single transformer decoder [that] generates reasoning tokens and action tokens in one unified autoregressive stream," and its open-source scripts let developers download the checkpoints and begin zero-shot inference deployment within minutes.[9]
In addition to VLA models, Galaxea develops World Action Models (WAM), which explicitly model how visual observations evolve under actions. The company views VLA and WAM as complementary technologies: VLA handles direct action generation, while WAM provides physical-world understanding, related to the broader concept of the world model.
Galaxea has published research on Fast-WAM, a WAM architecture that achieves 190-millisecond inference latency compared to the 800 milliseconds typical of conventional WAM approaches.[20] Fast-WAM retains video co-training during training but skips future prediction at test time, decoupling physical understanding from video generation. The Fast-WAM research was conducted using 60 hours of teleoperated demonstrations collected on the R1 Lite platform.[20]
Galaxea's technology strategy emphasizes real-world data collection over simulation-based approaches. CEO Gao Jiyang has stated that "at least for the next three years, reinforcement learning with simulators will not lead to the ultimate goal" for upper-limb manipulation tasks.[1] The company's data engine uses teleoperated demonstrations from fleets of R1 robots deployed in real environments, combined with imitation learning techniques.
The Galaxea Open-World Dataset has supported external research efforts as well, contributing to NVIDIA's EgoScale project and Ant Group's Lingbot-VLA. The company plans to build what it describes as the world's largest real-scene dataset using millions of hours of real-world data.[4]
Galaxea maintains an active open-source presence through its OpenGalaxea GitHub organization. Key repositories include:
The open-source strategy serves dual purposes: it lowers adoption barriers for research institutions considering Galaxea hardware, and it builds a developer ecosystem around the company's platforms. As of early 2026, over 150 developers use Galaxea's platforms, including teams at Stanford (Fei-Fei Li's group), Physical Intelligence, and NVIDIA.
Galaxea serves more than 40 clients across research, technology, automotive, electronics, and internet sectors.[4][7] Notable clients and deployments include:
| Sector | Clients |
|---|---|
| Technology and internet | Huawei Cloud, ByteDance, Physical Intelligence, NVIDIA |
| Automotive and manufacturing | Volkswagen, Haier |
| Electronics | Samsung |
| Academic and research | Stanford University, MIT |
| Government and public services | Beijing E-Town smart elderly care station (March 2026) |
Applications span algorithm training, robotic process automation, logistics (bin picking, line feeding, kitting, parcel handling), data collection for embodied AI, manufacturing automation, and service robotics. In the second half of 2025, Galaxea secured orders or letters of intent for thousands of units from companies in the automotive and logistics sectors.[19]
Gao Jiyang has outlined a three-stage development roadmap for the company:[1]
| Stage | Timeframe | Objective |
|---|---|---|
| Stage 1 | 3 to 5 years | Focus on one robot form factor and primary application scenario to achieve stable revenue |
| Stage 2 | 6 to 8 years | Expand to multiple robot forms addressing a broader range of scenarios |
| Stage 3 | Beyond 8 years | Develop a highly generalizable AI "brain" deployable across various hardware platforms, analogous to the evolution of GPT in language AI |
The company's guiding concept is "one brain, multiple forms," referring to unified visual perception and motion control foundation models that can eventually be deployed across different robot morphologies.[1] This includes developing both upper-limb manipulation models and lower-limb traversal and locomotion models.
Galaxea has announced plans to release a bipedal (two-legged) humanoid robot as part of its 2026 product roadmap, expanding beyond the wheeled R1 platform.[19] This would allow the company to address scenarios where legged locomotion is necessary, such as stairs, uneven terrain, and environments designed for human bipedal movement.
CEO Gao Jiyang has predicted that humanoid robots capable of cooking and cleaning will enter homes within less than a decade.[2] The G0 VLA models have been demonstrated on household tasks including cleaning, bed-making, and object retrieval. The company targets eventual consumer-facing deployment after establishing profitability through commercial and industrial applications.
The company aimed to ship up to 1,000 R1 units by the end of 2025, split between domestic and overseas markets including the United States.[2] For 2026, Galaxea plans to begin volume production at the 10,000-unit level as supply chain capabilities, particularly the partnership with Lens Technology for hardware and mass production, reach maturity.[19] The company was targeting profitability by 2026.
Galaxea Dynamics operates in a rapidly growing Chinese humanoid robotics market. By early 2026 it had reached unicorn status alongside Unitree Robotics, AgiBot, and Galbot, the four startups most often grouped at the front of China's embodied-intelligence sector.[3] According to the Beijing Academy of Artificial Intelligence, the competitive landscape in 2026 is organized into tiers:
| Tier | Companies |
|---|---|
| First tier | Unitree Robotics, AgiBot (Zhiyuan Robotics), UBTECH Robotics, Galbot |
| Second tier | Galaxea AI, LimX Dynamics, Xiaomi, Meituan |
| Third tier | Casbot, SenseTime, XPeng |
Compared to bipedal humanoid competitors like Unitree's H1 and G1 or UBTECH's Walker, Galaxea's wheeled R1 design offers advantages in stability and immediate deployability for indoor tasks, at the cost of terrain versatility. The R1's starting price of approximately $28,000 USD is competitive within the Chinese market, where several companies have pushed entry-level humanoid platforms below $30,000.[16] Galaxea also differs from arm-and-mobile-base specialists such as Galbot and from full-stack robot-arm makers like Fourier Intelligence by pairing its hardware with an open-sourced foundation model.
Galaxea differentiates itself from competitors through several factors:
CEO Gao Jiyang has predicted that 4 to 5 companies will coexist in the market as the industry matures, with success depending on technical understanding, scenario insight, and execution speed.[1]
Galaxea's registered legal entities include Xinghaitu (Beijing) Technology Co., Ltd. and Xinghaitu (Suzhou) Artificial Intelligence Technology Co., Ltd.[22] The Beijing office serves as the company headquarters, while the Suzhou facility at 88 Jinji Lake Avenue, Suzhou Industrial Park, houses additional operations. The "Galaxea Dynamics" brand name has been registered as a trademark with the United States Patent and Trademark Office.[21]
CFO Luo Tianqi has stated that over half of China's future top 10 embodied AI companies will emerge from the current generation of startups, reflecting the company's confidence in its positioning within the sector.[1]