This week I chose stories not about adding robots, but about deciding what the robots you have added should do.
Amazon is rebuilding the delivery station. It has spent more than a decade automating the warehouse itself: storing goods, gathering them to order, packing them into boxes. What remains is the step after that: sorting parcels by delivery route, arranging them in loading order, and handing them to the van. Amazon says a parcel in its European delivery stations passes through roughly 40 process steps between arrival and dispatch. These steps have resisted automation because what they handle changes daily. Parcel sizes, destinations, routes, vehicles and departure times are never the same twice. So the work has stayed manual.
According to internal planning documents, the new design would move parcels about 2.5 times faster than existing delivery stations. Amazon says the investment figures and timelines in those documents do not reflect current plans. The number matters less than the problem being solved. The question is not whether a robot can lift a parcel. It is where each of several thousand parcels should sit, when it should come out, and in what order — and how that order is rebuilt when the dispatch plan changes. That is not a machine problem. It is a decision problem.
Mike Harris, who leads the mobile robotics business at Ocado Group in the UK, makes the same point differently. What warehouses lack, he argues, is not technology but connected decisions. A robot picking faster achieves little if the pallet that item belongs on is still waiting at goods-in. What determines the result is whether picking, replenishment, returns, sorting and dispatch run as one sequence in which the outcome of each step feeds the instructions for the next, rather than as five separate operations under one roof.
Swisslog, the Swiss automation supplier, is talking about the stage before purchase, and I think that is the right place to start. Before choosing equipment, it says, you should have the numbers: your own historical order profiles, SKU velocity, inventory volumes and seasonal peaks. Design for an average day, Swisslog says, and you will get something that performs very differently once it runs your real operation. So work out, before you place the order, what happens when order composition shifts and when several sales channels have to be served at once.
Robot suppliers have also grown more candid about what their machines can and cannot do. Roy Belak, who leads robotic grasping at Locus Robotics, says suction is all you need 60–70% of the time. The hard part is the remaining 30–40%. And what should be measured is not picks per hour but the quality of the pick: whether the item was damaged, whether two were taken at once, whether the wrong one was taken. None of this appears on a spec sheet, but it reaches the cost line quickly.
Prices, meanwhile, are falling. JPMorgan estimates the cost of running a humanoid is dropping below $10 an hour, against roughly $30 an hour for a warehouse worker. Whether you can buy one is now a separate question. The US FCC added foreign-made humanoid and quadruped robots to its Covered List, halting new imports. About 85% of the global humanoid market is Chinese-made.
So robots get cheaper, while the list of countries you may buy them from shifts with politics. Equipment selection and procurement will stay hard to predict. One thing does not move with outside events: the decision layer — the software that tracks what is waiting where in your own facility, then tells machines from different vendors, and the people working alongside them, what to do next and where to send it. Add machines without it and, as Ocado puts it, you simply move the bottleneck: pick faster, and the goods wait at the next step instead. What Amazon is trying to solve at the delivery station, what Ocado calls connected decisions, and what Swisslog says you must settle before choosing equipment are all the same layer.
GROUND built GWES with that decision layer at the centre of the product, rather than assembling it later from other vendors’ software. Japan’s logistics facilities face the world’s tightest labour shortage, handle very wide SKU ranges, and run several different work flows through the same space. That is why we believe productivity turns on having it before adding machines.
GROUND Inc. — Founder & CEOTomo Miyata
This week
Five stories
01Lead storyAt the delivery station, parcels are sorted by route, arranged in loading order and handed to the van. Photo: Unsplash
Logistics OS & decisionsUnited States
Amazon — “Project Tetromino”: letting machines decide where each parcel sits and in what order it comes out
Amazon is developing a new delivery-station design known internally as Project Tetromino, as reported by Business Insider and analysed by Logistics Viewpoints. A delivery station receives parcels that have passed through the fulfilment centre and line haul, sorts them by delivery route, sequences them, and hands them to drivers. Amazon says a parcel in its European delivery stations passes through roughly 40 process steps between arrival and dispatch. The company has invested more than €700 million in technology at its European delivery stations.
On the warehouse side, AI, robotics and automation have been arriving for more than a decade. Delivery stations lagged because what they handle never settles: parcel sizes, destinations, routes, vehicles and departure times change daily. Internal planning documents indicate the new design would move parcels about 2.5 times faster than existing delivery stations, and set out an investment plan including a $103 million initial pilot. Amazon told Business Insider that Tetromino is an early-stage concept and that the figures and timelines in the documents do not reflect current plans.
The problem is not whether a robot can lift a parcel. It is where each of several thousand differently sized parcels should sit right now, when it should come out, and in what order it should be fed to the loading position across hundreds of delivery routes — and how that order should be rebuilt when the dispatch plan changes. One technology Amazon is evaluating comes from Boxbot, which pairs automated storage with software for sorting, sequencing and retrieval. Boxbot claims vehicle loading up to ten times faster (its own figure, not an Amazon projection). FedEx is working on the same problem, handling irregular parcels, with Dexterity AI.
Ocado — “Warehouses don’t need more tech. They need connected decisions”
This is a contributed piece by Mike Harris, who leads the mobile robotics business at Ocado Group in the UK. Ocado sells the automation technology it developed while running its own online grocery operation. Harris argues that mobile robots have already proved they cut walking time and raise picking productivity, but that this is only part of the problem. Orders, labour, congestion and exceptions cannot be optimised separately. Picking, replenishment, returns, sorting and dispatch need to run as one sequence in which the outcome of each step feeds the instructions for the next, not as five separate operations under one roof.
While decisions stay split across systems and roles, automation simply moves the bottleneck. A robot picking faster achieves little if the pallet that item belongs on is still waiting at goods-in, or if the exception it created is still sitting in someone’s work queue. Ocado’s Ocado IQ continuously evaluates order urgency, pick locations, available labour, workload and congestion, then regroups work, re-prioritises and reassigns as conditions change. Rather than telling a robot where to go next, it decides what the operation should do next.
The company also splits its fleet into three tiers: base units that cover daily volume, extra units for routine peaks such as Monday start-ups or promotions, and seasonal units for peaks you can plan for, such as Black Friday. Ocado says its mobile robots deliver two to three times the productivity of trolley picking, cut labour costs by up to 50%, can go live in as little as 14 weeks, and pay back in six months. They are deployed at more than 120 sites.
Swisslog — “Design from your own operational data before you choose the equipment”
A position piece from Swisslog, the Swiss logistics automation supplier, aimed at customers in Australia and New Zealand. Part of the KUKA group, Swisslog implements automated storage systems including AutoStore and supplies SynQ, its integration and orchestration software. Steve Dimitrovski, its head of sales, argues that automation is a major investment and should not be judged only on what it can do on day one.
What should inform the design, he says, is your own historical order profiles, SKU velocity, inventory volumes, seasonal peaks and how requirements are likely to change. “Design for an average day and you will end up with something that performs very differently in real operation.” You need to understand what happens as volumes grow, as order composition shifts, and when several sales channels must be served at once. He also recommends simulating to find bottlenecks before ordering equipment, and designing expansion in from the start rather than bolting it on later.
He makes a point about comparing bids as well: two proposals can look similar on paper while the engineering behind them differs substantially. The questions to ask are what data were used, under what conditions the design was validated, what the binding constraints are, and how the system can be extended later.
Locus Robotics — suction covers 60–70%; measure pick quality, not speed
Locus Robotics has acquired Canada’s Nexera Robotics and will fit Nexera’s soft gripping technology, NeuraGrasp, to Array, the Locus mobile picking robot. Locus builds mobile robots for warehouses; its fleet has completed more than 6 billion picks. Array, announced by Locus in March 2025, is a mobile robot with a vision-guided arm. Nexera had been through at least six product generations before the acquisition. Array, separately, already had the air and power supply NeuraGrasp needs.
Roy Belak, formerly Nexera’s CEO, now leads robotic grasping at Locus. Robotic manipulation, he says, sits on a spectrum between suction and pinch grasping. “Suction is very good, particularly in warehouse operations. Sixty to 70% of the time, it’s all you need. It’s when you get to that 30% to 40% and get to the really high coverage ratios.” NeuraGrasp uses both suction and pinching. Belak expects pinch grasping to become central, but says reliability falls as complexity rises.
Speed is explicitly not the priority. “Picking has historically been a field where the claims have been more focused on the return on investment, without understanding the underlying quality. We can do many picks per hour, but what people don’t focus on is how good are those picks.” What to watch, he says, is product damage, double picks and mispicks — none of which appear on a spec sheet, all of which reach the customer’s cost line quickly. The obstacles he still sees are the intelligence needed to judge a pinch grasp, and touch sensing robust enough not to degrade with use. Simulated data, in his view, will not solve touch.
JPMorgan — humanoid running costs fall below $10 an hour, just as the US halts foreign imports
JPMorgan expects demand for humanoid robots in US manufacturing to rise sharply. As reported by MarketWatch, the cost of running a humanoid for an hour is falling below $10, against roughly $30 an hour for a warehouse worker. The bank sees US manufacturing, where vacancies go unfilled, as the first market.
Procurement conditions, however, have changed. The US Federal Communications Commission (FCC) added foreign-made humanoid and quadruped robots to its Covered List of restricted equipment, halting new imports on national security grounds. The target is China, which accounts for about 85% of the global humanoid market. Chinese foreign ministry spokesperson Mao Ning said protectionism would not make the US more competitive and would only harm American companies and consumers, adding that China would take all necessary measures.
The shipment numbers are still small. Research firm Omdia puts 2025 global humanoid shipments at about 15,000 units. China’s Unitree and AGIBOT each shipped more than 5,000; Tesla and Figure AI each shipped a few hundred. Unitree listed on Shanghai’s STAR Market this month, closing its first day 460% above the offer price at a valuation of roughly $50 billion, after raising about 6.1 billion yuan ($905 million). Its 2025 revenue was 1.7 billion yuan, up from 393 million yuan a year earlier, with net profit of 278 million yuan; about 44% of core business revenue came from outside China.
