Galbot G1

The humanoid that gave up on legs and got a job

Galbot’s G1 is 1.73 m tall with a 1.9 m arm span, and it does not walk. It rolls, on a 360-degree omnidirectional wheeled chassis with a foldable-leg design that lets the torso rise and drop across a 0 to 2.4 m vertical working range. No bipedal gait, no backflips, no viral parkour clip. What it does have is a claimed grasping success rate above 95% on randomly stacked, previously unseen objects — including the two categories that break most vision systems, transparent and specular.

That is the whole strategy in one sentence. Galbot, founded in Beijing in May 2023 by Peking University’s Wang He, bet that the commercially valuable hard problem in embodied AI is not locomotion but manipulation, and that the fastest way to get manipulation data is to put robots in real stores and let them work.

170 stores, and why that number is misleading

By WAIC in July 2026, Galbot was operating more than 170 “Galbot Space Capsule” unmanned stores across 40-plus cities, plus over 100 smart-pharmacy instant-retail warehouses, some running autonomously for over a year. One robot covers a 50 m² store carrying roughly 5,000 SKUs, and a site can be brought up in a day. The target is 1,000 stores within three years.

Be precise about what this is. These stores are operated by Galbot, as Galbot’s own retail business. They are not customers buying robots. Anyone citing “170 deployments” as commercial traction is counting a company’s own shops. The honest traction numbers are elsewhere: 1,200+ G1 units cumulatively shipped in 2025, a JD.com listing at RMB 630,000 live since February 2026, and industrial orders from CATL, Bosch, Toyota, SAIC, BAIC, Zeekr, Great Wall and Hyundai.

Even so, the stores matter enormously — just not as revenue. They are the largest continuously-operating grasping dataset in a real commercial environment that anyone has assembled. Every bag of chips retrieved is a labelled trial.

Synthetic data first, real data second

Galbot’s models invert the usual order. GraspVLA, published in January 2025, was pretrained on SynGrasp-1B — one billion frames of synthetic action data generated from over 10,000 object meshes across 240 Objaverse categories — then post-trained on real data. GroceryVLA handles the retail case specifically: soft bags, boxes, bottles, glass, without per-SKU tuning. TrackVLA does vision-only navigation and following. The umbrella is AstraBrain, fed by an AstraSynth synthetic-data pipeline. Galbot was also the first company to put NVIDIA’s Jetson Thor into a humanoid.

The company’s claim of “1,000× Tesla’s training efficiency” is corporate marketing and we are not repeating it as fact. The published SynGrasp-1B methodology, however, is real and peer-reviewed at CoRL 2025.

What we could not verify

Galbot’s own site is JavaScript-rendered and does not expose a machine-readable spec sheet, so several numbers in circulation come only from aggregators and contradict each other: total DOF is variously 21 or 47; weight is 80, 85, or 92.5 kg; runtime is 8 or 10 hours; the end effectors are described both as a suction cup plus parallel gripper and as 12-DOF dexterous hands. We have published only the figures traceable to Galbot’s own exhibitor filing and left the rest out rather than pick a plausible-looking number. If you see a Galbot G1 spec table quoting max speed of 1.5 m/s, that traces to a single SEO blog that labels it “reported.”

One more correction worth carrying: reporting that “nearly 300 units” sold on JD.com during Chinese New Year is wrong. That figure was added to cart. Global Times reported that just two units were listed and sold out instantly.

Why we care for LostJobs

The convenience-store clerk is not usually on anyone’s list of jobs threatened by humanoid robots, because the popular image of the threat is a biped in a car factory. Galbot has been quietly demonstrating the opposite: the first job a humanoid takes end-to-end is more likely to be retail than manufacturing, because a shop is a flat-floored, well-lit, climate-controlled environment with a finite object set and a customer who does not care who fetched the item.

At RMB 630,000 the machine does not yet beat a clerk on pure cost in most of the world. But Galbot is not selling against a clerk’s wage; it is selling a store that runs at 3 a.m. with no staff at all. That is a different arithmetic, and it is already winning in 40 Chinese cities. Watch the CATL production-line results — if the same grasping policy holds up in light assembly, the addressable job pool jumps by an order of magnitude.