Past midnight in a Beijing pharmacy, a wheeled robot navigates the cramped aisles alone. Its dual arms reach for medications on cluttered shelves—ibuprofen here, cough syrup there—fulfilling orders for customers who will collect their purchases in the morning without ever entering the store. The machine, built by a startup called Galbot, can supposedly identify roughly 5,000 different products, pick them from messy environments, dodge obstacles, and complete an entire fulfillment cycle without human help.
The company says it's deployed this system—the G1, they call it—in over a dozen Beijing pharmacies and aims for 100 by year's end, all running around the clock.
Perhaps more than the founders expected, those pharmacy robots represent something quietly seismic for retail. This isn't another shelf-scanning robot that cruises grocery aisles counting inventory. It's a machine that actually manipulates objects in live retail environments, trained on what researchers call vision-language-action models—VLA for short—that fuse computer vision, natural language understanding, and motor control into a single neural architecture. The technical shift here: moving from narrow perception tasks to what the industry calls "embodied intelligence." Systems that can generalize across products, store layouts, and instructions without needing to be reprogrammed every time someone rearranges an endcap or introduces a new SKU.
If it works—and that's still an if—the implications ripple well beyond pharmacies.
When Robots Start Touching Things
Retailers have spent years deploying perception robots. Simbe Robotics' Tally robot scans shelves in roughly 1,000 U.S. grocery stores as of 2024. Brain Corp's floor cleaners scrub aisles autonomously. But manipulation—actually picking things up and moving them—has remained confined to the back of house, in controlled warehouse zones where lighting is consistent, shelves are standardized, and humans stay out of the way.
Galbot's live deployments, however preliminary, signal that the technical barriers are eroding faster than many anticipated. VLA-driven robots that can reliably pick and place items in dense, public-facing environments change the calculus around front-of-house automation entirely.
The technical foundation matters. VLA models collapse what used to be separate robot subsystems—vision, language parsing, path planning, motor control—into a unified neural network. Google DeepMind demonstrated the core concept with RT-2, published in 2023: a vision-language model trained on both internet images and robot trajectory data could output action commands directly from visual and text inputs. RT-2 improved out-of-distribution generalization from 32% (its predecessor RT-1) to 62% and could chain together reasoning steps before executing movements.
Then came Stanford's OpenVLA in 2024, a 7-billion-parameter open-source model trained on the Open-X-Embodiment dataset—over a million robot trajectories collected across different robot bodies and tasks. It matched or exceeded RT-2's performance on some manipulation benchmarks, proving that open models could reach competitive capabilities if fed sufficient cross-embodiment data.
Galbot's approach builds on this lineage but targets retail-specific scenarios. The company's research group at Peking University published GraspVLA in May 2025, a grasping foundation model trained on SynGrasp-1B—a billion-frame synthetic dataset combining autoregressive perception with what researchers call flow-matching action generation. The model handles open-vocabulary grasping, meaning it can pick objects it's never seen in training if given a natural-language description.
A second model, TrackVLA, addresses visual tracking and navigation. It uses a shared language-model backbone with separate heads for object recognition and trajectory planning, trained on EVT-Bench's 1.7 million samples. An October 2025 extension, TrackVLA++, added spatial reasoning and target memory for complex tracking tasks in spaces like malls or hospitals, where crowds shift unpredictably.
These aren't purely academic exercises. Galbot founder and CTO He Wang holds a joint appointment at PKU and runs the PKU-Galbot Joint Lab of Embodied AI. The lab's work feeds directly into the company's commercial platform, shortening the loop between research and deployment in a way that mirrors Covariant's trajectory in warehouse robotics. Whether that translates to retail dominance is another question entirely.
The Pharmacy Gambit—And the Numbers Behind It
Galbot launched in May 2023 and raised approximately $335 million by July 2025, according to industry reports. The company later claimed over $800 million in funding and a valuation near $3 billion by December 2025. Those figures come from company press releases, though, and lack independent verification—a detail worth noting in an ecosystem where promotional excess sometimes runs ahead of operational reality.
The funding environment reflects broader enthusiasm. Goldman Sachs projected a $38 billion humanoid robot market by 2035. Morgan Stanley's long-range forecast reaches $5 trillion by 2050, albeit with caveats about cost compression and reliability improvements that haven't materialized yet.
The G1 robot itself stands 173 cm tall, a wheeled semi-humanoid with dual arms rated for 5-kilogram payloads and a 10-hour runtime. Galbot claims the robot's body folds for transport and that deployment in a new store takes one day. During a Beijing robotics shopping festival, the company priced units at 680,000 RMB—roughly $93,000. The pharmacy use case involves receiving digital orders, locating medications across thousands of SKUs, picking them from shelves, packaging, and either dispensing at a kiosk or staging for delivery. Unmanned, round-the-clock operation, the company says.
Beyond pharmacies, Galbot has announced partnerships that suggest industrial ambitions. A joint venture with Bosch's investment arm Boyuan Capital focuses on embodied AI for smart manufacturing. According to a June 2025 press release, the JV aims to build "an end-to-end value chain" for globally competitive solutions—corporate speak, but it signals intent. The company also claims thousands of industrial unit orders from manufacturers including CATL, Toyota, and Hyundai. Those figures remain unaudited.
A separate JV with Youlife targets vocational training for robot operation, an acknowledgment that even advanced AI systems need human oversight in deployment.
Galbot's hardware runs on NVIDIA's Jetson Thor platform, part of NVIDIA's broader push into what it calls "physical AI." NVIDIA released its Isaac GR00T N1 foundation model in 2025—an open architecture for humanoid robots that partners including Boston Dynamics and Agility Robotics have adopted. Galbot's integration positions it to benefit from shared model improvements and synthetic data pipelines. It also means the company competes in an increasingly crowded field of Thor-enabled platforms, where differentiation depends less on hardware and more on domain-specific data and deployment speed.
The Data Race Nobody's Talking About Enough
Galbot isn't working in a vacuum, of course.
Covariant, a California-based robotics AI company, launched RFM-1 in 2024—an 8-billion-parameter foundation model trained on tens of millions of robot-action trajectories from deployed warehouse systems. Covariant's model handles manipulation tasks like bin picking and item sorting across diverse objects. The company claims it collects a million new trajectories every few weeks from its fleet. That data flywheel—deploy robots, collect edge cases, retrain models, push updates—gives Covariant a structural advantage in controlled warehouse environments that's hard to replicate without scale.
NVIDIA's GR00T initiative takes a different angle: provide an open foundation model and synthetic data tooling so that any partner can fine-tune for their embodiment and domain. The February 2026 update introduced GR00T N1.6 alongside Cosmos Reason, a world-modeling architecture that generates synthetic training data from video. This approach democratizes access to large-scale pretraining but also commoditizes it. Whoever has the best domain-specific data and fastest deployment cycle wins.
The academic community continues pushing boundaries. DexGraspVLA, published in early 2025, tackles dexterous grasping with multi-fingered hands using hierarchical VLAs. Google DeepMind has hinted at on-device versions of its Gemini-based robotics models to reduce latency and connectivity dependence—critical for retail environments where network reliability can be inconsistent. The Open-X-Embodiment dataset now includes augmentations bringing total trajectories above 5.4 million. EmbodiedScan adds a million egocentric views for first-person navigation training.
All of this points to convergence. VLA models are getting large enough and generalist enough that sim-to-real transfer works reliably for constrained tasks. The barrier to deploying a new manipulation skill is dropping from months of engineering to days of fine-tuning.
At least in theory.
Why Aisles Aren't Warehouses
Despite the technical progress, front-of-house retail manipulation remains operationally harder than warehouse picking. Much harder, actually.
Warehouses are structured environments: fixed racks, controlled lighting, predictable object placement, minimal human traffic. Retail floors are chaotic. Shoppers move unpredictably. Endcaps change weekly. Products get shoved back in the wrong spot. Lighting varies by aisle. A toddler runs down the cereal section while someone's service dog investigates the pet food endcap.
Walmart discontinued its contract with Bossa Nova in 2020 after deploying shelf-scanning robots across stores. The official line cited shopper experience concerns and cost-benefit calculations. The robots were good at perception but still required human intervention for interpretation and restocking decisions. Translation: they didn't close the ROI loop.
Manipulation adds another layer of complexity—and liability. A robot arm operating near customers triggers safety questions and must comply with collaborative standards like ISO/TS 15066, which governs power and force limits. Illinois BIPA lawsuits targeting in-store biometric and computer vision systems—such as a 2024 case against Target that was allowed to proceed—underscore privacy sensitivities. California's CPPA finalized rules in 2025 requiring cybersecurity audits, risk assessments, and transparency for automated decision-making systems, with compliance windows extending into 2026-2030. Any VLA system processing shopper movements or biometric proxies must navigate these frameworks.
The lawyers, in other words, haven't finished weighing in.
Simbe Robotics, which has arguably achieved the most scaled deployment in U.S. grocery with Tally robots in approximately 1,000 stores, focuses exclusively on perception and data capture. CEO Brad Bogolea described shelf intelligence as "foundational infrastructure" at Groceryshop 2025, noting board-level mandates driving adoption. Simbe's model involves remote operations—humans review the data Tally collects—and the robot never touches products.
Consumer sentiment studies Simbe commissioned (400 shoppers) showed 77-80% positive or neutral attitudes toward in-store robots. But that acceptance is predicated on unobtrusive, non-contact operation. The moment a robot arm reaches across the aisle near your shopping cart, sentiment probably shifts.
Galbot's pharmacy deployments sidestep some of these issues by operating in environments with no or minimal customer presence during fulfillment windows. The stores function as dark fulfillment centers accessible via digital orders, not traditional browse-and-buy formats. Smart wedge strategy. But it doesn't prove the robots can handle mixed-traffic grocery aisles during peak hours—which is where the real market opportunity lives.
Other players are taking similarly cautious approaches. Amazon piloted Agility Robotics' Digit humanoid in fulfillment centers for tasks like moving totes, not picking individual SKUs in public-facing stores. Apptronik, which raised $935 million in a Series A extension through February 2026, targets logistics and back-of-house operations with its Apollo robot. Figure AI's partnership with BMW focuses on factory automation.
Common thread: manipulation in controlled spaces first. Then expand to less structured environments as models improve and regulatory clarity emerges—if it emerges.
The China Acceleration (and the Bubble Warnings)
China's robotics ecosystem is moving at a different pace, fueled by provincial subsidies and national policy priorities around embodied intelligence. Galbot benefits from this environment directly. Beijing's robot shopping festival offered subsidies up to 250,000 RMB per robot purchase, effectively cutting the G1's price tag by more than a third. UBTECH announced plans for mass production of its Walker S2 humanoid and claimed large orders from enterprises. Unitree shipped low-cost humanoid platforms like the G1 (a different robot—confusingly sharing Galbot's model name) aimed at developers and researchers.
However.
China's National Development and Reform Commission issued a warning in November 2025 about an emerging humanoid investment bubble, noting overcapacity risks and speculative valuations. Galbot's claimed $3 billion valuation and reports of "thousands of units ordered" exist in a disclosure environment with limited third-party auditing. The company's assertions should be treated as such until independent metrics surface.
This tension—genuine technical progress alongside promotional excess—isn't unique to China, to be fair. Apptronik's >$5 billion valuation and Galbot's funding claims reflect capital markets pricing in aggressive growth scenarios that may not materialize. The difference is that Western deployments tend to happen in partnership with large anchor customers (Amazon, BMW, Mercedes) who impose operational discipline. Some Chinese deployments are driven by government procurement and subsidies that can evaporate if policy priorities shift.
Which they do, in China. Frequently.
What the Next Two Years Look Like
The near-term trajectory splits along environment type.
Pharmacies, convenience stores, micro-fulfillment centers, and back-of-house retail zones offer constrained settings where VLA-driven manipulation can prove out economically. Expect more pilots in these formats over the next 18-24 months. Walmart's partnership with Symbotic—automating all 42 regional distribution centers and building new high-tech perishable DCs—shows where scaled automation investment is flowing: the supply chain, not the sales floor. By early 2026, Walmart projects roughly two-thirds of stores will be serviced by automated distribution, with unit costs down 20% and 55% of fulfillment center volume automated.
Front-of-house manipulation in high-traffic grocery remains aspirational. Robots will likely handle specific tasks—restocking after hours, staged BOPIS order assembly in back rooms, inventory verification with light repositioning—before they pick and place items in aisles during shopping hours. The VLA models are getting good enough. But the surrounding infrastructure (safety certification, liability frameworks, real-time replanning in crowded spaces) lags behind.
One wildcard: on-device model deployment.
If VLA models shrink enough to run entirely on robot hardware without cloud dependencies, latency and connectivity concerns evaporate. Google DeepMind's work on on-device Gemini robotics variants and NVIDIA's push for edge inference on Jetson platforms both point in this direction. Retailers prioritize data sovereignty and network reliability. Robots that don't need continuous cloud connectivity have an operational edge, particularly in stores with spotty wifi or where data privacy concerns run high.
The data advantage matters more as the technology commoditizes. Covariant's million trajectories per few weeks from warehouse deployments create a feedback loop that academic datasets can't match. Galbot's pharmacy network—if it scales to hundreds of stores as projected—could generate similarly rich domain-specific data. The question is whether retail manipulation data proves as valuable as warehouse data, given the higher variability and lower task repetition in retail environments. Warehouses are about volume. Retail is about variety.
Goldman Sachs' $38 billion humanoid market projection for 2035 and Morgan Stanley's $5 trillion long-term forecast hinge on cost compression—actuators, hands, batteries—and on VLA models becoming reliable enough to justify the capital expense. Aging demographics and structural labor shortages drive demand, certainly. But only if robots achieve 95%+ uptime and handle edge cases without constant human babysitting.
Kroger's retrenchment from Ocado automated fulfillment centers in 2025, closing multiple CFCs amid a shift to store-fulfilled delivery, is a reminder that automation ROI can be elusive even with mature technologies. Sometimes the low-tech solution—having store employees pick orders between stocking shifts—simply pencils out better.
The Test Case
For retail executives evaluating these systems, the calculus is straightforward but unforgiving.
Scanning and data-capture robots like Simbe's Tally and Focal Systems' ceiling cameras (300,000+ cameras, 2 billion+ labeled images) have crossed into proven infrastructure. The business case is legible, the vendors are known, and the deployments number in the hundreds or thousands. Manipulation robots operating in public-facing environments are still crossing the chasm. The technical foundations appear solid. The question is whether the next 24 months bring deployments that prove the unit economics work outside of controlled pilots and government-subsidized showcases.
Galbot's pharmacy deployments are a test case worth watching. If the company hits its 100-store target and the robots maintain reliable uptime, it validates the VLA approach for at least one retail format. If the rollout stalls or unit economics don't close, it signals that the gap between lab results and operational reality remains wider than the funding frenzy suggests.
Either way, the midnight pharmacy in Beijing—robot arms reaching across cluttered shelves, no human in sight—offers a glimpse of what retail's future might look like. Assuming, of course, the robots keep working when the subsidies run out and the venture capital moves on to the next big thing.
The industry will know soon enough.
