Galbot G1
Last reviewed
Sources
33 citations
Review status
Source-backed
Revision
v2 · 4,353 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
33 citations
Review status
Source-backed
Revision
v2 · 4,353 words
Add missing citations, update stale details, or suggest a clearer explanation.
Galbot G1 is a wheeled, dual-arm mobile manipulator built by Beijing Galbot Co., Ltd., a Chinese embodied-AI start-up that also goes by the name Galaxy General (Chinese: 银河通用). The robot pairs a human-like upper body, two arms, and a foldable torso lift on top of an omnidirectional wheeled base, and is the centerpiece of the company's pitch for a general-purpose robot that can run a roughly 50-square-meter convenience store on its own. By the end of 2025 Galbot had raised more than $800 million in cumulative venture funding across three years, reached a $3 billion valuation after a December 2025 round of over $300 million, and put fleets of G1 units into more than 30 Chinese cities running 24/7 unmanned retail and pharmacy locations.[5][3][7]
Galbot sits in a slightly different corner of the humanoid market from bipedal contenders like the Agility Robotics Digit, Figure 03, or 1X Neo. The company has been explicit that wheels are not a stopgap but a deliberate design choice for indoor commercial tasks, where a wheeled base trades the spectacle of walking on two legs for longer runtime, lower cost, and far better reliability on flat floors. In that respect the G1 is closer in spirit to Reflex Robotics than to the bipedal field, and the founder Wang He has used the wheeled-versus-legged debate as a recurring talking point in interviews.
The Galbot G1 is a semi-humanoid mobile manipulator: a wheeled robot with two dexterous arms and a vertically extending torso, rather than a bipedal walker. Galbot describes it as a semi-humanoid that combines a human-scale upper body with two highly dexterous arms on a 360-degree omnidirectional wheeled base.[2][15] It is designed for embodied AI tasks on flat indoor floors, picking, restocking, retrieving, and handing over goods in convenience stores, pharmacies, warehouses, and factory aisles. The G1 was first shown publicly in mid 2024 and moved into commercial production through 2025.[4]
Beijing Galbot was registered on 19 May 2023 in the Zhongguancun Science Park, Haidian District, Beijing.[32] The founding team came out of the embodied AI research lab at Peking University, where Wang He was running a group on dexterous manipulation and three-dimensional vision.[31] The company name in Chinese is 银河通用机器人, which translates literally as "Galaxy General Robotics," and the English brand was eventually shortened to Galbot. In several early news cycles the firm was also called GalaxyBot before settling on the current name.[11]
Galbot was set up to combine a full-stack embodied-AI software pipeline with a robot it could sell into commercial deployments. The company has framed itself as a vertically integrated player that builds its own data pipeline, foundation model, and hardware in house, describing itself as "the world's first company to fully achieve full-stack in-house development across hundreds-of-billion-scale high-quality datasets, embodied foundation models, and robotic hardware."[5]
Wang He (Chinese: 王鹤) is an assistant professor at Peking University and the founder and chief technology officer of Galbot. He earned a bachelor's degree from Peking University and a PhD from Stanford University, where he worked on three-dimensional vision and robot learning.[31] After Stanford he returned to Peking University as a faculty member, set up an embodied AI lab inside the School of Computer Science, and became a BAAI Scholar at the Beijing Academy of Artificial Intelligence.[31] His research group has published widely on dexterous manipulation, simulated grasping, and category-level pose estimation.
In May 2024 Wang formalized the company-university link by setting up the Peking University, Galbot Joint Laboratory of Embodied Intelligence, with himself as director. The lab channels academic research into Galbot's product pipeline and provides a recruiting funnel. Galbot has also established a research center with BAAI and a clinical robotics partnership with Xuanwu Hospital in Beijing.[3]
Wang has been a regular speaker at the World Robot Conference in Beijing and at the World Humanoid Robot Games, where Galbot took the top prize in the robot skills competition in August 2025.[14]
Galbot is headquartered in Beijing with additional research and development sites in Shenzhen, Suzhou, and Hong Kong.[5] The Beijing office is colocated with the Peking University joint lab. Shenzhen handles manufacturing engineering, Suzhou is positioned near the automotive customer base in the Yangtze River Delta, and the Hong Kong office is used for international hiring and overseas partnerships. By late 2025 Galbot reported a headcount of several hundred employees across the four sites and described itself as a vertically integrated robotics company that develops its dataset pipeline, foundation models, and robot hardware in house.[5]
The Galbot G1 is a wheeled mobile manipulator with two arms, a multi-segment vertical lift, and a sensor head. It is not bipedal and does not walk. The decision to drop legs was deliberate: Galbot's commercial use cases are flat indoor floors in convenience stores, pharmacies, warehouses, and factory aisles, where wheels are faster, safer, and cheaper than a bipedal gait.[4]
Galbot publishes a subset of the G1 specifications in marketing materials, and third-party trackers such as humanoidintel.ai, aparobot.com, and humanoid.guide have collated additional figures.[15][16][17] The values in the table below are drawn from those sources. Where individual outlets disagree the table reflects the most-cited published figure.
| Item | Galbot G1 specification |
|---|---|
| Robot type | Wheeled bimanual mobile manipulator |
| Height (extended) | About 1.73 m |
| Reach (top) | Up to 2.4 m with torso lift extended |
| Reach (bottom) | Down to floor level with torso lowered |
| Mass | About 55 to 88 kg, varies by configuration |
| Total degrees of freedom | 47 |
| Arm configuration | Two arms, roughly 7 DoF each |
| Dexterous hand | 12 DoF per hand, multi-finger |
| Mobile base | Omnidirectional wheeled platform |
| Torso lift | Vertical stroke of approximately 65 cm |
| Maximum payload | About 5 to 20 kg depending on task and arm pose |
| Battery runtime | About 4 to 8 hours per charge in deployed use |
| Top travel speed | Around 5 km/h on flat ground |
| Power | Lithium battery, swappable in newer units |
| Sensors | Stereo cameras, depth sensors, tactile fingertips |
| Compute | On-board GPU plus separate motion controller |
The G1's most distinctive hardware element is the vertical lift in the torso. The robot can compress to about 1 meter tall to reach low shelves and floor-level objects, and extend the lift to operate near a 2.4 meter top shelf without bending.[15] This is a key feature for convenience-store work, where typical shelving runs from the ground up to two meters and conventional fixed-height humanoids would need to crouch or stretch awkwardly.
Each arm has roughly seven degrees of freedom and is designed for bimanual cooperation rather than single-arm dexterity. The hands have twelve degrees of freedom each, putting them at the dexterous end of the current humanoid market where many competitors ship with five or six DoF grippers.[15] Galbot uses a mix of cable-driven and motor-driven finger actuation, with tactile sensing in the fingertips. The extra dexterity supports tasks in retail and pharmacy operations, including picking individual items off densely packed shelves, manipulating small consumer-goods packaging, and handing items to a human customer.
The base is an omnidirectional wheeled platform with what Galbot describes as zero-radius turning, similar in concept to a four-mecanum-wheel mobile robot.[2] The combination of an omnidirectional base, a torso lift, and dexterous arms gives the G1 a large workspace per footprint, which is the key reason a 50 square meter unmanned convenience store can be operated by a single robot. The wheeled choice has trade-offs: the G1 cannot climb stairs, cannot handle outdoor terrain, and is not intended for outdoor environments. In exchange it gets several extra hours of battery life than comparable bipedal humanoids, a lower total cost of ownership, and a simpler safety case for operation around the public, since the robot cannot fall over the way a bipedal robot can.
The sensor stack combines stereo cameras around the head, depth cameras on the torso, force and tactile sensors in the hands and arms, and inertial sensors in the base. The compute pipeline runs perception, planning, and low-level control on board, with optional cloud connection for fleet coordination and remote monitoring. Galbot has not published a full bill of materials but its VLA model literature implies the G1 carries a discrete GPU with enough memory to run distilled vision-language models locally.[33]
Galbot's software approach is centered on large vision-language-action (VLA) models trained on synthetic data, with the fine-tuned policies then transferred onto the G1 hardware. The company has released several named models that together make up the embodied-AI stack:
The company's training pipeline is heavily synthetic. GraspVLA is pre-trained on SynGrasp-1B, a billion-frame synthetic grasping dataset built from over 10,000 object meshes spanning 240 categories with photorealistic rendering and extensive domain randomization, with real-world teleoperation data used only for fine-tuning and corner cases.[33] The GraspVLA paper describes the model as "a VLA model pretrained on large-scale synthetic action data as a foundational model for grasping tasks" that shows "advanced zero-shot generalizability and few-shot adaptability" while helping to "mitigate sim-to-real gaps."[33] The motivation is both scale and coverage: simulation generates orders of magnitude more interaction data than teleoperation, and can sweep through object variations and physical parameters that would be expensive to stage in the real world.
This is a different bet from the predominantly teleoperation-driven data pipelines at 1X Technologies and Figure AI. Galbot's argument is that physics-grade simulation with strong sim-to-real transfer scales further than human teleoperators ever can, and that the same model can be deployed across heterogeneous robot bodies, with the TrackVLA demonstration on Unitree's H1 offered as evidence. Galbot supplements the foundation models with a model-based controller layer for low-level motion, force control, and safety interlocks. The hybrid architecture is similar in shape to Agility Robotics' approach on the Digit but leans more heavily on learned policies for manipulation.
Galbot's flagship commercial product is the Galbot Store, a fully autonomous convenience store run by a single G1 robot. The format is a roughly 50 square meter unmanned shop with about 5,000 stock-keeping units stored in 6,000 shelf slots, operated 24 hours a day with no in-store staff.[8] Customers walk in, scan items on a self-checkout, or order through a mobile app and have the robot retrieve products from behind a counter. Galbot has said a single G1 can be deployed to set up a store in about one day.[4]
The first Galbot Store pilots opened in Beijing in 2025 with about ten locations. By late 2025 the company had expanded the format across more than 30 Chinese cities, with deployments reaching tens of locations in major markets including Beijing, Shanghai, and Shenzhen.[7] Galbot projected approximately RMB 100 million (about $14 million) in retail revenue for 2025 from the store network.[8]
The stores are a critical proof point because they show full robotic operation in front of paying customers rather than a controlled industrial demo. The robot handles inventory checks, restocking from a back room, item retrieval, packaging, and basic customer interaction.[4] Galbot has shared per-store SKU counts and restock cycles, but has not published detailed throughput or unit-economics figures.
Galbot has expanded the same wheeled-manipulator format into 24/7 unmanned pharmacies. By late 2025 the company was operating close to ten pharmacies in Beijing with a stated plan to scale to 100 locations across Beijing, Shanghai, and Shenzhen by year end.[4][14] The pharmacies use the same G1 hardware running a pharmacy-tuned policy that handles drug-package picking, prescription verification, and customer hand-off through a window.
In the hospital setting Galbot partnered with Beijing's Xuanwu Hospital to test G1 units in patient rooms, in-hospital pharmacies, and guidance and triage roles.[3] The hospital partnership is structured as a clinical-grade pilot rather than a paid commercial deployment, and was designed both to validate the safety case for the G1 in close human contact and to generate data on hospital-specific tasks.
Galbot has also pushed into industrial manipulation through proof-of-concept projects with major automakers and battery manufacturers.[12] Disclosed customers and partners include:
| Customer | Sector | Role of G1 |
|---|---|---|
| CATL | EV battery manufacturing | Material handling and assembly support, pilot |
| Bosch | Auto parts and industrial automation | Joint venture for automotive applications |
| Toyota | Automotive | POC manufacturing-line trials |
| Hyundai | Automotive | POC manufacturing-line trials |
| Mercedes-Benz | Automotive | Scanning, plate sorting, bin transport, POC |
| Zeekr | Electric vehicles | Material handling, POC |
| BAIC Group | Automotive | Production line trials |
| SAIC | Automotive | Production line trials |
| Great Wall Motor | Automotive | Production line trials |
The industrial program is structured differently from the retail one. Where the convenience-store fleet runs Galbot's own retail operation, the automotive deployments are integrator-led pilots on customer manufacturing floors. Galbot has reported orders for thousands of units across the manufacturing pipeline, though specific volume figures for individual customers have not been disclosed.[3] The company has framed itself as the first to deploy humanoid-style robots for real autonomous operations on a car factory floor, a claim contested by Figure AI's BMW Spartanburg fleet and Apptronik's Mercedes-Benz pilot.[5]
In June 2025 Galbot signed a strategic joint venture with Boyuan Capital, the investment arm of Bosch Group, and Bosch China.[6] The venture, called the Boyin Innovation Alliance, targets embodied AI for automotive and industrial applications. Dr. Ingo Ramesohl, head of Bosch Ventures, stressed the role of embodied AI in next-generation manufacturing in announcements.[6] The tie-up provides a credibility anchor in Western industrial markets and a route into European factory pilots.
Beyond retail and manufacturing, Galbot has run a warehouse-automation program with what the company describes as stable 24/7 operations for more than a year. The deployments use the G1 for tote handling, shelf restocking, and cross-dock transport. Galbot has not published per-customer throughput numbers, so the logistics figures are less directly comparable to the published GXO numbers from Agility Robotics' Digit fleet.
Galbot has raised more than $800 million in cumulative venture funding since its founding in 2023, reaching a $3 billion valuation after a December 2025 round of over $300 million.[5][7] The funding has come from both Chinese venture capital and large strategic investors. The table below lists the publicly disclosed rounds.
| Round | Date | Amount | Lead and notable investors | Valuation context |
|---|---|---|---|---|
| Angel | 2024 (mid year) | About CNY 700 million (about $96 to 100 million) | Meituan, BAIC Industrial Investment, SenseTime Guoxiang Fund, iFlytek Fund, Qiming Venture Partners, Lanchi Ventures, Matrix Partners China, Source Code Capital, IDG Capital | Record-breaking angel round for Chinese embodied AI |
| Strategic | July 2025 | About $151 to 153 million (CNY 1.1 billion) | CATL (via Xiamen Puquan Fund), Puquan Capital, China Development Bank, STIF, Beijing Robot Industry Fund, Qiming Venture Partners | Cumulative funding above $330 million |
| Series unspecified | December 2025 | Over $300 million | Investors from China, Singapore, and the Middle East | $3 billion post-money valuation, cumulative roughly $800 million |
| Follow-on | Early 2026 | About CNY 2.5 billion (about $363 million) | National AI Fund and follow-on participants | Valuation reported above $3 billion |
The July 2025 strategic round of RMB 1.1 billion (about $153 million) was led by CATL and Puquan Capital, with China Development Bank, the Beijing Robot Industry Fund, and Qiming Venture Partners participating; it brought cumulative funding to about $335 million at the time.[4][10][26] The December 2025 round of over $300 million set what the company called records for both the largest single-round financing and the largest cumulative financing in China's embodied AI sector.[5]
Cumulative reported funding through early 2026 puts Galbot among the best-capitalized humanoid robot companies globally. Its $3 billion valuation as of December 2025 is comparable to Figure AI's mid-2024 valuation and higher than Agility Robotics' Series C valuation of about $2.12 billion. Chinese-language coverage from outlets such as Gasgoo has described the period from mid 2025 through early 2026 as Galbot raising close to CNY 5 billion in three months at a valuation above CNY 20 billion.[29][30]
The investor list is unusual for a Chinese robotics company. Meituan, the on-demand services giant, was an anchor in the angel round, which gave Galbot a strategic link to one of China's biggest urban-services networks.[9] CATL came in through its private equity vehicle in 2025, tying the company to the largest EV battery maker in the world.[10] The Bosch venture arm joined through the joint-venture structure rather than direct equity, but is functionally a strategic investor. The mix is heavy on industrial strategics rather than pure financial venture funds, mirroring the company's commercial pipeline.
The humanoid market in 2026 is split between several technical archetypes. Galbot G1 is the most visible commercial example of the wheeled-base archetype, with Reflex Robotics' wheeled humanoid as the closest peer in the United States and Apptronik's Apollo and Agility's Digit representing the bipedal alternative. The table below shows a side-by-side comparison of the main contenders.
| Robot | Maker | Country | Locomotion | Approx. height | Total DoF | Hand DoF | Primary deployments |
|---|---|---|---|---|---|---|---|
| Galbot G1 | Beijing Galbot | China | Wheeled, dual arm | ~1.73 m | 47 | 12 per hand | Unmanned stores, pharmacies, factory pilots |
| Reflex humanoid | Reflex Robotics | United States | Wheeled, dual arm | Variable via lift | Not disclosed | 3-finger gripper | Warehouse, food service |
| Digit | Agility Robotics | United States | Bipedal | ~1.75 m | 24 | 2-finger gripper | GXO RaaS contract, Amazon pilots |
| Apollo | Apptronik | United States | Bipedal | ~1.73 m | 28 to 30 | 5-finger | Mercedes-Benz, Jabil, GXO pilots |
| Figure 03 | Figure AI | United States | Bipedal | ~1.68 m | 30+ | 5-finger | BMW Spartanburg fleet, home pilots |
| Optimus Gen 3 | Tesla | United States | Bipedal | ~1.73 m | 28+ | 22 DoF hand | Tesla internal logistics |
| Neo | 1X Technologies | Norway / US | Bipedal | ~1.65 m | Not disclosed | 5-finger | Home pilots |
| Unitree H1 | Unitree | China | Bipedal | ~1.80 m | 27 | Optional 6 DoF | Research, demos, TrackVLA tests |
Galbot's twelve DoF hand puts it at the high end of dexterity for current commercial humanoids, ahead of Reflex's three-finger gripper and Digit's two-finger gripper.[15] Its 47 total DoF is higher than most peers because it counts hands and torso joints alongside the arms and base. On the operational side, Galbot is the only company in the wheeled-base camp running its own end-customer retail operation rather than selling robots to a third-party logistics provider. Agility Robotics has shipped more cumulative paid-customer throughput through the GXO contract, while Figure AI has more visibility at BMW.
Reflex Robotics is the closest direct peer in design philosophy. Both companies sell a wheeled dual-arm manipulator with a vertical lift, both target indoor commercial environments, and both argue against bipedal locomotion as a near-term commercial bet. Reflex is a small US start-up focused on warehouse and food-service work, while Galbot has a far larger operation, deeper VLA model investment, and a captive retail deployment network in China. Among Chinese peers, Galbot's wheeled, manipulation-first approach contrasts with the bipedal locomotion focus of Unitree Robotics and the broad humanoid programs of companies such as Galaxea Dynamics.
Galbot has generated heavy coverage in Chinese tech press through 2024 and 2025, with a steadily rising profile in international robotics media. The Robot Report, IEEE Spectrum, Robotics and Automation News, and TechCrunch have all covered the company's funding rounds and deployment milestones.[3][4][6] South China Morning Post, 36Kr, and Caixin have followed the company in Chinese-language coverage, often framing it alongside competitors like Unitree, Fourier, and UBTECH as the leaders of China's humanoid push.[30]
In August 2025 Galbot won gold in the robot skills competition at the inaugural World Humanoid Robot Games in Beijing, an event hosted by the Beijing government and the China Electronics Standardization Institute.[14][28] In the pharmaceutical-sorting final the G1 ran fully autonomously, without the teleoperation most of the 22 competing teams relied on, finishing with 336 points (about 160 points ahead of the second-place team) and a time of 10 minutes and 22 seconds.[14] The win was reported in Chinese state media as evidence of the country's progress in embodied AI, and Wang He used the platform for a high-profile keynote on the wheeled-versus-legged debate.[28]
Not all coverage has been positive. Skeptics have argued that the unmanned-store format is a constrained environment that disguises rather than tests the G1's general manipulation ability. Store SKUs can be carefully selected to match the robot's strengths and the back-of-house restocking still relies on human-loaded carts. Others have noted that the published revenue projections are aggressive given the small number of stores. The Robot Report has been sympathetic on the technology but cautious on the financial projections.[3]
A second criticism is about cross-embodiment claims. Galbot's TrackVLA demonstration on the Unitree H1 was widely covered, but some observers questioned how much of the policy transfers to a bipedal robot versus the wheeled G1. On the labor side, the unmanned-store deployments have drawn limited criticism in Chinese media about retail-job displacement, and the pharmacy deployments have drawn closer regulatory scrutiny because of the higher stakes of drug dispensing.