Picture a convenience store in Zhongguancun that closes for no one. At 3 a.m., a wheeled humanoid behind the counter reheats a sausage, boxes it, and slides it toward a pickup shelf. No camera crew, no crowd — just a shift that never ends.
Most of the world's attention on Chinese humanoids goes to bipedal sprints and gymnastics. Galbot (银河通用) made a quieter bet: skip the legs, keep the arms, and put the machine to work on a schedule humans cannot match.
Key takeaways
- Galbot's G1 is a wheeled, dual-arm humanoid built around AstraBrain (银河星脑), an embodied large model the company describes as integrating "brain, cerebellum, and neural control" end to end.
- The company says its "Galaxy Space Capsule" (银河太空舱) retail format has landed in more than 40 Chinese cities across 170-plus sites.
- More than 100 instant-retail warehouses now run Galbot units on 7×24 duty, a streak the company says has passed one and a half years.
- Industrial deployments are documented at battery maker CATL and at Bosch and Toyota lines, handling inspection, transport, and sorting.
The body: a cart with arms
Galbot G1 stands up to 1.73 m tall and weighs 92.5 kg. It rides a four-wheel omnidirectional base at up to 1.5 m/s rather than walking, a choice that trades the spectacle of bipedal motion for uptime and stability. The arms offer 21 degrees of freedom (two in the neck, fourteen across both arms, three at the waist, two at the legs), with a single-arm load of 5 kg and a combined 10 kg. Compute comes from an NVIDIA AGX Orin with 64 GB and 275 TOPS. The unit is rated IP54 and runs about 8 hours on a charge that takes 3.5 hours.
End effectors swap between a dexterous hand, a suction cup, and a gripper, which lets one platform move between picking, placing, and suction tasks without a hardware swap. That modularity is the practical heart of the "general-purpose" claim.
The brain: closed-loop control
Galbot markets AstraBrain as the differentiator. Instead of separating perception, planning, and control into stacked modules, the company says its system links multimodal perception with real-time feedback control so the robot "perceives, understands, decides, and executes" in a loop. The pitch is resilience: in a dynamic store or warehouse, the machine keeps operating as conditions change rather than freezing when a plan breaks.
Where the machines actually are
Galbot's most concrete footprint is retail and warehousing. On its own product page, the company states the Galaxy Space Capsule format has reached more than 40 cities and over 170 sites, and that more than 100 instant-retail warehouses run Galbot units around the clock, with continuous operation reportedly past eighteen months. In an August 2026 China Today feature, co-founder Zhang Zhizheng said the robots independently run smart retail spaces in over 100 stores across more than 20 Chinese cities.
Industrial use is narrower but name-brand. Zhang told China Today the robots have entered production lines at CATL, Bosch, and Toyota, taking on automated quality inspection, transport, and parts sorting, and that the company has partnered with several top hospitals for medication sorting and ward delivery. Those statements are the co-founder's own and were reported, not independently audited.
The two retail footprint figures come from different sources and scopes: the "40+ cities / 170+ sites" number is the company's own product-page claim, while the "over 100 stores / 20+ cities" figure comes from the co-founder's China Today interview, so they should not be read as conflicting measurements of the same thing.
Why wheels, not feet
Bipedal humanoids grab headlines because they look like us. They also fall, tire, and cost more to keep upright. Galbot's wager is that the near-term value of a humanoid is the upper body — the dexterous arms and the brain — not the legs. A wheeled base reaches the same shelf height, never loses balance, and runs a triple shift.
That logic fits China's retail and logistics density, where stores and micro-warehouses sit close to customers and labor costs for overnight and repetitive handling keep climbing. A machine that simply shows up every night has a clearer payback than one that learns to run.
According to Li An, chief scientist at BrainNet (脑机网), the first durable robot wages will likely be won not by the most human-looking machine, but by the one whose uptime and task mix make the economics obvious to a store manager.
Li An's comments reflect his own views and do not represent NeuroAI's editorial position or an endorsement by BrainNet (脑机网).
Honest limitations
Figures for store counts, city coverage, and warehouse uptime come from Galbot's own marketing pages and from a co-founder interview in China Today; none are independently audited. The CATL, Bosch, and Toyota references are attributed statements, not confirmed deployment disclosures by those companies. "Embodied large model" and "brain-cerebellum-neural control" are the company's framing, not a standardized technical category, and we have not tested task success rates or safety in person. Robot density per site, maintenance cost, and failure modes are not public.
What readers can do now
- Look for Galbot's 7×24 warehouse claim to be matched by a third-party operator case study — round-the-clock duty is where uptime promises get tested.
- Watch whether CATL, Bosch, or Toyota reference Galbot in their own automation communications, which would convert a founder quote into verified demand.
- Compare Galbot's wheeled approach with bipedal rivals on total cost per task hour, not on demo videos.