Banned in America, Dispensing in Berlin: Galbot’s G1 Robot Works Real Pharmacy Shifts at IFA

IFA Berlin opened September 4 at Messe Berlin, and the most consequential exhibit in Hall 25 may not be the most famous brand on the floor. Galbot — a three-year-old, Beijing-headquartered company that has quietly become the highest-valued unlisted humanoid robotics firm in China — is operating Booth 160 in IFA Next, where visitors can watch a Galbot G1 robot retrieve items from shelves, navigate a crowded hall, and hand objects to strangers with no operator in the loop.

The demonstration is a live rehearsal of something the company already does at scale at home. In Beijing’s Haidian district, a licensed pharmacy — the first in China to receive a pharmaceutical business license for robot-assisted dispensing — runs Galbot robots handling 5,000 SKUs at a reported 99.5% pick success rate. More than 170 autonomous Galbot Store retail units now operate across 40-plus Chinese cities, including a FamilyMart location in Beijing’s Zhongguancun district where the G1 autonomously takes orders, retrieves products, and hands them off.

What most visitors are unlikely to see posted on the booth wall: 38 days before IFA opened, on July 28, 2026, the US Federal Communications Commission added foreign-produced advanced robotic devices to its Covered List — formally classifying robots in exactly the G1’s hardware category as posing “an unacceptable risk to US national security” and barring new models from the equipment authorization required for importation, marketing, or sale in the United States. Galbot’s IFA debut is, in part, an explicit response: locked out of the American market, the company is pushing into Europe, where no equivalent ban exists.

What the G1 actually is

The G1 is not the bipedal robot science fiction trained audiences to expect. It is a wheeled semi-humanoid mobile manipulator — a design that deliberately trades legs for an omnidirectional wheeled base to gain stability, energy efficiency, and operational continuity on the flat commercial floors where it is actually deployed.

At 173 centimeters tall with an arm span of 190 centimeters and a working reach of 240 centimeters, the G1 is built to operate inside spaces designed for people, without the balance constraints that make bipedal robots hard to commercialize. Each arm provides seven degrees of freedom and ends in a 12-degree-of-freedom dexterous hand, with a 5-kilogram payload per arm (10 kg combined). The robot runs up to 10 hours per charge — a full commercial shift — and carries 47 total degrees of freedom. Compute is an NVIDIA Jetson AGX Orin 64GB module in the standard configuration, with Jetson Thor available in the G1 Premium.

The sensor suite reads like a security review checklist: a head binocular camera, four torso RGB cameras, wrist depth cameras, force sensors, chassis LiDAR, eight ultrasonic sensors, and cloud connectivity for AI updates and remote monitoring. That last item is precisely what the FCC flagged.

AstraBrain and the VLA stack

The technical story under the commerce is a training architecture. Galbot calls its integrated AI system AstraBrain, described internally as a “brain-cerebellum-neural control” architecture, running three specialized Vision-Language-Action (VLA) models on the same hardware:

• GraspVLA handles object manipulation — the end-to-end model that turns camera feeds, natural-language instructions, and joint positions directly into motor commands.
• GroceryVLA handles retail task sequencing.
• TrackVLA handles navigation and target-following, letting the robot follow a human through a dynamic environment and recover after line-of-sight interruptions, without LiDAR-based pre-mapping of each site.

A VLA model eliminates the traditional separation between perception, planning, and control software. Operationally, this means nobody writes motion trajectories for each new object: when a supplier changes a package size, the operator gives a new language instruction instead of reprogramming the robot. On top of the three models sits a world model that simulates the outcome of a candidate action before executing it — the difference between a robot that reacts and one that plans.

The zero-shot generalization that makes this work is trained in NVIDIA Isaac Sim at a scale no real-world data collection could match: billions of simulated physical interactions generate the training signal for GraspVLA before a physical robot touches a real object. The “reality gap” is addressed through domain randomization — intentionally varying simulation parameters so the model generalizes rather than memorizes. Galbot reports 95–97% grasping success across deformable packaging, glass bottles, and rigid containers (company-stated figures, not independently audited), and claims one-day deployment into a new retail location.

The factory proof point

If the pharmacy is the showcase, the industrial contract is the anchor. Galbot’s S1 industrial robots have run 24 hours a day at CATL’s Ningde battery production base since passing acceptance testing in March 2026, under a procurement contract worth CNY 236 million (about $35 million). The company counts procurement orders from CATL, Bosch, Toyota, Hyundai, and BAIC, with deployments in German facilities through partners including a Boyuan Capital joint venture. In February 2026, two G1 units listed on JD.com at roughly CNY 630,000 (about $94,000) sold out in minutes after a Spring Festival Gala appearance before an estimated 679 million television viewers.

Galbot was founded in May 2023 by Professor He Wang of Peking University, who leads a joint laboratory between the company and the university. A CNY 2.5 billion (about $362 million) March 2026 round brought total funding to roughly $800–900 million at a $3 billion post-money valuation, with investors including China’s National AI Industry Investment Fund, Sinopec, CITIC, Bank of China Asset Management, SAIC, CATL, Meituan, Sequoia China, IDG Capital, and Qiming. The company is reportedly exploring a Hong Kong listing at a valuation of up to $4 billion — a path Unitree cleared first with its August 19 STAR Market debut in Shanghai.

Why the FCC ruling matters in Berlin

The Covered List mechanism — first applied to Huawei in 2021, then DJI drones in December 2025, then foreign-made consumer routers in March 2026 — now extends to mobile robots weighing more than 4.4 pounds that combine environment-perceiving sensors, wireless connectivity of at least 200 kilobits per second, and software controlling autonomous navigation, perception, data collection, or remote command-and-control. The G1 meets all three prongs by a wide margin. The FCC identified four risk categories: manipulation of collected data or physical operations; surveillance of US persons or sensitive locations; enhancement of foreign intelligence capabilities; and unauthorized remote access or commandeering of the robot.

For European buyers in Hall 25, the FCC’s determination is legally irrelevant — and that is the story’s sharpest edge. The EU has no equivalent list for Chinese-origin robots, and no EU regulation currently requires Galbot to disclose what data its robots collect or how it is handled. China’s National Intelligence Law (2017), Article 7, requires Chinese organizations and citizens to “support, assist and cooperate” with state intelligence work on demand — an obligation that follows the manufacturer regardless of where the robot is deployed, what a privacy policy says, or where data servers sit. A G1 in a German factory or Dutch warehouse continuously records visual, spatial, and operational data about that facility, and no procurement contract can remove the underlying legal condition. No publicly documented independent security audit of the Galbot G1 exists as of IFA’s opening day.

The mitigations available are technical, not legal: network segmentation isolating the robot from production systems, firmware audits before deployment, disabling cloud update endpoints. They remove the pathway, not the obligation.

Two markets, one robot

Galbot is not alone in Hall 25 — AGIBOT occupies Booth H25-371 and EngineAI is at H25-349 — but it arrives with the most commercially validated story: real pharmacy licenses, real 24-hour factory shifts, real consumer demand at $94,000 a unit. The ET1, its first bipedal humanoid unveiled at the World Robot Conference in August, played autonomous tennis against former world No. 15 Zheng Jie with an approximately 90% forehand contact rate on balls above 50 km/h — a research platform pointing at household tasks.

The bifurcation now hardening across the robotics industry is the real headline from IFA’s Hall 25. A Chinese embodied-AI company with verified commercial deployments, state-linked capital, and a $4 billion IPO trajectory is barred from the world’s largest economy by administrative ruling — and is marketing directly to the world’s second-largest Western market, which has drafted no equivalent response. European industrial buyers get capable robots with one-day deployment and 99.5% pick rates, plus a legal exposure profile their regulators have not yet decided how to address. Whether Berlin marks the beginning of European uptake or the moment the EU starts drafting its own covered list is the question IFA 2026 leaves open.