Period covered 7 – 13 September 2026Written by Tomo Miyata
This week I chose three stories about how many robots to install, and two about what matters once they are in.
JD Logistics, the logistics arm of the Chinese e-commerce group JD.com, said on 9 September that it will procure 3 million robots, 1 million unstaffed delivery vehicles and 100,000 delivery drones over five years. What I noted was not the number but the software shown the same day, Meta Brain 3.0. It brings storage, transport and delivery decisions into one system and has cut route calculation for hundreds of millions of parcels from minutes to seconds. The 3 million robots are being procured on the assumption that this software will direct them.
Numbers are rising in Europe too. Hai Robotics of Shenzhen, a maker of rack-climbing robots (ACR), is installing more than 1,500 in a new facility for a European fashion retailer. Arthur Bellamy, chief revenue officer of Exotec of France, expects tighter US import authorisation for foreign-made robots to push Chinese manufacturers toward Europe, and warns of subsidised, low-priced products entering the market. Robots on their own will keep getting cheaper and more plentiful. This week's three stories support that view.
So what matters once they are in? I chose a piece in Supply Chain Management Review by Neal McGuckin, who has run operations at Amazon's Middle East and North Africa business and at IAG Cargo in the UK. Amazon acquired Kiva Systems in 2012, and in 2025 its robot fleet passed one million. DeepFleet, introduced that year, is an AI that sequences and routes the fleet; it improved travel efficiency by about 10 per cent without adding a single machine. The biggest constraint left after a million robots was not machine performance but coordination between machines.
His conclusion is that automation has an order: standardise totes and pallets so work can be measured, automate transport around that standard, mechanise low-variability processes one at a time, and only then install the software that directs the whole. Buy equipment first and goods pile up between machines while exceptions exceed what supervisors can handle. Unless a process runs by hand given correct information before it is automated, its defects are locked in. He calls automating an unfixed process "automating the mess".
That order corresponds to the five stages of GPIM, the method GROUND uses to diagnose the operational maturity of a logistics facility. Stage 1, Systematisation, aligns procedures and units of work; Stage 2, Visualisation, puts progress, results, inventory and forecast workload into numbers. Stage 3, Automation, moves routine work to machines and systems; Stage 4, Optimisation, decides inventory placement, movement, staffing and delivery with AI and mathematical optimisation. Stage 5, Autonomy, is the whole running without waiting for human decisions. His standardisation is Systematisation, his information discipline is Visualisation, his process-by-process mechanisation is Automation, and DeepFleet is Optimisation. The failure he describes is buying Stage 3 and 4 equipment while skipping Stages 1 and 2. GROUND takes the proper steps in order before proposing any equipment, so that no stage is skipped.
Buyers' criteria are changing as well. Maven Robotics of the United States, founded in 2024 by former Apple engineer Hamza Derbas, raised $100 million on 10 September. Its robot, two arms on a wheeled base, builds store-specific mixed pallets. In a selection run by a large consumer-goods company, it won over four competitors that already had products. What Derbas explained was not performance but the scope of the job: "We're not trying to solve a single robot problem. It hooks in to a warehouse management system on one side; product goes on trucks on the other." Eight robots now run 16 hours a day at 99 per cent uptime or better.
Apply this to logistics facilities in Japan. Robots, components and training data will keep getting cheaper. Two things will separate operators: whether one piece of software can tell machines from different manufacturers, and the people working alongside them, what to do in what order and where to send things, and whether people can draw the line on how much the AI decides. Add machines without software that issues those instructions, and the congestion and stoppages described above follow. GROUND's GWES, its software for logistics facilities, puts that instruction-issuing function at the centre of the product rather than bolting it on to third-party software. In Japan, where labour shortages are severe and high-variety goods flow through the same building, I believe the presence or absence of that capability will show up directly as a difference in productivity.
GROUND Inc. — Founder & CEOTomo Miyata
This Week's News
Five stories
01Lead storyJD Logistics will procure 3 million robots over five years and bring storage, transport and delivery decisions into one piece of software. Photo: Unsplash
Physical AILogistics OS & OrchestrationChina
JD Logistics: 3 million robots in five years, with storage, transport and delivery decisions in one piece of software
At JDDiscovery 2026, the technology event JD.com held in Beijing on 9 September, its logistics subsidiary JD Logistics unveiled a robot series called Wolf. According to Cao Peng, chairman of JD.com's technology committee, the company will procure 3 million robots, 1 million unstaffed vehicles and 100,000 delivery drones over five years. Wolf picks, sorts, transports and delivers with minimal human involvement, and includes units for minus 20°C environments and automated pharmacy dispensers. Meta Brain 3.0, shown the same day, brings storage, transport and delivery decisions together and cut route calculation for hundreds of millions of parcels from minutes to seconds.
JD Cloud will collect more than 10 million hours of video from real operations over two years as training data, and is building a 100,000-chip computing cluster on domestic silicon with GPU designer Moore Threads. JD Property will open more than 80 RoboBase hubs covering research, pilot production and maintenance within five years, and JD Industrials aims to become the world's largest supplier of robot components. JD.com and its affiliates plan to retrain their roughly 700,000 delivery and logistics workers for technical roles such as robot maintenance.
CommentaryLogistics OS & OrchestrationUnited States
The order Amazon's million robots reveal: the last gain came from software directing the fleet
A commentary by Neal McGuckin in Supply Chain Management Review. McGuckin has run operations at Amazon's Middle East and North Africa business and at the British air-cargo carrier IAG Cargo. Amazon acquired Kiva Systems in 2012, and in 2025 its deployed fleet passed one million robots, close to its workforce of 1.2 million.
He identifies three failure modes: specifying equipment against unmeasured processes; automating the movement of goods when the constraint is informational; and automating everything at once, creating a structure in which one stoppage halts the whole. Amazon standardised the unit of work before automating, then added machines process by process for sorting, carton lifting and cart transport. DeepFleet, introduced in 2025, is an AI that sequences and routes the fleet; it improved travel efficiency by about 10 per cent with no new equipment. His sequence is standardisation, transport automation, process-by-process mechanisation, then the software that directs the whole. Unless a process runs by hand given correct information, automation locks in its defects.
Hai Robotics: 1,500 rack-climbing robots for a single European fashion facility
Hai Robotics of Shenzhen makes autonomous case-handling robots (ACR) that climb racking to retrieve cases. On 9 September it announced it will supply more than 1,500 HaiPick Climb robots to a new e-commerce facility for a leading European fashion retailer, which it did not name. It is the largest deployment of the type to date.
The facility covers 30,000 square metres, with 1.2 million double-deep storage locations and throughput above 24,000 totes an hour. All 1,500 robots run under one control system, and the equipment fits standard existing racking. "This project demonstrates the continued shift toward highly scalable robotic automation within large-scale fulfilment operations," said Peter Guan, general manager of Hai Robotics EMEA.
Maven Robotics raises $100 million after winning a deal on scope of work, not on a robot
Maven Robotics disclosed a $100 million round on 10 September from RoboStrategy, LocalGlobe, Vine Ventures and XTX Ventures. Co-founder and chief executive Hamza Derbas spent nine years in Apple's special projects group and founded Maven with his brother Khalid when the group was disbanded in 2024. The robot has two arms on a wheeled base and handles loads of up to 30 kilograms.
Its main task is mixed palletising: rebuilding single-product pallets from factories into store-specific pallets, work that in most facilities is still done by people walking the floor. In a selection run by a large consumer-goods company, Derbas presented the scope of the job rather than the robot's performance: "We're not trying to solve a single robot problem. It hooks in to a warehouse management system on one side; product goes on trucks on the other." Maven won over four competitors that already had products. Two years on, eight robots run 16 hours a day at 99 per cent uptime or better. The funding will build 250 third-generation units. Derbas rejects two-legged designs as "very complex, unreliable" and costly.
Regulation & Market StructureEurope & United States
Exotec: US rules on foreign-made robots may steer Chinese manufacturers toward Europe
The US Federal Communications Commission (FCC) has tightened authorisation for imports of foreign-made robots. European systems already authorised in the United States are not automatically affected, but companies selling into the US will be asked where systems are made, who supplies the components and who owns the company, and will be assessed against new security requirements.
Exotec of France makes automated storage systems built around the rack-climbing robot Skypod. Arthur Bellamy, its chief revenue officer, says the FCC framework "is not designed to target individual countries. European companies should not assume that their origin alone provides an exemption." In the earlier FCC measures on drones, companies from Israel, Austria and Germany received conditional approvals; no Chinese company did. He expects Europe to become a relatively more attractive market for Chinese manufacturers. His concern is price: products backed by state support and preferential financing entering Europe at prices European companies cannot match. Some European brands already sell automation developed and manufactured by Chinese companies. Europe, he argues, should add scrutiny of subsidies, ownership and product origin to its safety and cybersecurity requirements.
