None of the five stories I chose this week is about the robots themselves.
STILL, a German maker of forklifts and logistics automation, has set out what a warehouse control system does. A WCS is the software that tells the machines inside a facility which load to move next, and where to send it. STILL calls it the digital nerve centre of automated warehouse logistics. If you are running a few AGVs, the vendor’s own fleet software is enough. But once conveyors and shuttle storage enter the same flow of goods, no single vendor’s software can direct every machine. Loads then wait between machines, and work stops. So a WCS should not be added after the equipment has been bought. It belongs in the tender requirements from the start.
Infios, a major supplier of supply chain execution software, has set out a three-stage approach to how much AI is allowed to decide. Suppose bad weather delays an inbound shipment by half a day. Do you re-sort the load, re-assign the labour, or wait? The question is who makes that call. At the first stage, AI proposes and people decide whether to act. At the second, people set the rule in advance: handle projected delays of up to 48 hours automatically, and report anything beyond that for human supervision. At the third, AI weighs the size of the delay within that rule and tells people only when it needs to. At every stage, the reason for the decision is recorded. Whether AI can be trusted to act depends less on the accuracy of its forecasts than on how far people have drawn the line, and whether the reasoning can be checked afterwards.
HireArt, an American company that supplies the people who run robots on the floor, points to a different problem. While five or ten robots are running in a controlled setting, a small team is enough. But when sites grow to several dozen, operations move to two- and three-shift working, and each floor has its own layout and travel paths, the nature of the work changes. It stops being the work of commissioning robots and becomes the work of keeping many sites running every day. At that point, measuring throughput speed alone is counterproductive: it rewards the operator who skips steps and pushes work through without reporting faults.
Robots themselves, meanwhile, keep getting easier to buy. LG Electronics and NVIDIA will bring online this year a facility built for one purpose: to have robots perform tasks and capture their motion as data. The target is 100,000 hours, and that data will train the AI that drives the robots. Schaeffler, the German bearings group, has changed how it makes the gearboxes used in humanoid joints, from machining to forming. Manufacturing cost falls by more than 25%, and the company begins volume production in 2027.
So capable robots and inexpensive components will both be available to buy. Once that is true, results on the floor are not decided by the choice of machine. They are decided by two things. The first is whether a single piece of software can tell the machines from different vendors, and the people working alongside them, what to move next and where to send it. The second is whether people have drawn a clear line on how much AI is allowed to decide. Add machines without that arrangement and, as STILL writes, loads wait between them and stoppages increase rather than fall.
GROUND’s GWES puts the software that directs the machines at the centre of the product. It is not other companies’ software joined together after the fact. Japan’s logistics facilities are the most short-handed in the world. They handle a wide range of items, and several streams of work run through the same space. That is why I believe that whether a company has this capability decides its productivity.
GROUND Inc. — Founder & CEOTomo Miyata
This week’s stories
Five stories
01Lead storyA WCS is the software that tells the machines inside a facility which load to move next, and where to send it. Photo: Unsplash
WCS & logistics OSGermany / Europe
STILL — “A modern WCS is the digital nerve centre of automated warehouse logistics”
STILL, the German maker of forklifts and logistics automation systems, has set out how software divides the work inside an automated logistics facility. Its parent, the KION Group, is one of the world’s largest industrial truck manufacturers and also owns Dematic, the WMS and automation systems business. STILL divides the software into three layers. The WMS tracks what is stored where, and how much of it. The WES decides the order in which incoming orders are broken down and assigned to work. The WCS issues instructions directly to the machines. The WCS plans routes, identifies faulty items, prevents collisions, and judges whether to stop or run. The company calls it the “pacesetter” of automated material flows, the element that sets the speed at which goods move.
The data volumes are large. A shuttle installation sends several hundred status and position messages for a single instruction to move a load. A fleet of AGVs exchanges position data with central software continuously. AGVs on their own can run on their vendor’s fleet software. But the moment a machine that software cannot address joins the same flow, a WCS is required. Bringing AGVs from different vendors under one controller means using higher-level software built to VDA 5050, the common communication standard drawn up by the German automotive industry association and others. Even then, STILL writes, the approach stops coping as unit counts and data volumes rise.
Combining WMS, WES and WCS reduces the number of seams between systems. With a single home for the data, the state of the operation is easier to see, and the number of systems and suppliers involved falls. “A modern WCS is the digital nerve centre of automated warehouse logistics. It orchestrates the interaction between conveyor systems, robotics and people in real time,” says Johannes Funke, who leads sales and engineering for automated intralogistics solutions at STILL.
Infios — “Graduated Autonomy”: three stages for how much AI decides
An interview with Infios, the supply chain execution software company formed by the German industrial group Körber and the American investment firm KKR. It supplies combined warehouse and transport management software to more than 5,000 companies in 70 countries. The interviewee is Eugene Amigud, its head of innovation and a 25-year veteran of supply chain and retail technology. He argues that a lack of trust is a key reason AI projects fail. A decision cannot be handed to a system the floor does not believe in. So you start small and widen what is delegated by degrees. Sense the situation, decide, act, learn from the result, sense again. Infios calls this way of building trust through repetition Graduated Autonomy.
At the Assisted stage, AI only proposes; people decide whether to act. At the Automated stage, people set the criteria in advance: the system might handle all projected delays of up to 48 hours, with exceptions beyond that reported for human supervision. At the Autonomous stage, AI weighs the size of the delay within those criteria and decides how to escalate as well. Of analytics dashboards, Amigud says: “They look great, they are clean and smart, easy to use. They show a lot of information, but the crucial difference is that they do not act on that information. Their usefulness is limited.” A record is kept at every stage: why the decision was made, which conditions were weighed and which were not.
“Robots don’t run themselves” — add sites and the workforce is the first constraint
An analysis published by The Robot Report. The author is Christopher Bower, co-founder and president of HireArt, a New York-based firm that handles contract hiring and employment on behalf of its clients and supplies the people who run robots on the floor. When a deployment moves from pilot to full rollout, he writes, what stalls is not the robot. It is the organisation that operates it, maintains it and keeps improving how it is run. While there are only five or ten robots and the environment is controlled, a few engineers and experienced operators are enough. But when sites grow to several dozen, shifts multiply, and each floor has its own layout and travel paths, the nature of the work changes. It is no longer product commissioning. It becomes the work of keeping many sites running, every day.
The roles emerging are robot operators, field maintenance technicians, remote operators, motion validators and data collectors. Each sits between engineering and the floor, working out why something unexpected happened, recording the failure, and reporting what happens on the floor back to the people who build the product. Measuring speed alone is counterproductive here: it rewards the operator who skips steps and pushes work through without reporting faults. What deserves measuring instead, Bower argues, is whether procedure was followed, whether the records kept are of good quality, whether the judgement to escalate a fault was correct, and whether the operator erred on the side of safety when unsure.
LG Electronics and NVIDIA — a facility to capture 100,000 hours of robot motion data, live this year
LG Electronics will bring its Robot Data Factory in Yangjae, Seoul, into full operation this year. It is a facility built for one purpose: to have robots perform tasks and capture their motion as data. The plan is to gather 100,000 hours within the same year, a volume equivalent to almost twelve years of continuous operation. Inside are rooms that replicate a home and a zone that reproduces work from the company’s washing machine plant in Tennessee. There is also a zone fitted with the logistics automation equipment sold by LG CNS, the group’s IT services arm, and a zone for testing robot hands from LG Innotek, its electronic components business.
The company’s home robot, CLOiD, works in each zone and generates the data. The data collected is then augmented with NVIDIA’s technology, which synthesises the missing scenarios, before it goes into training. The aim is to train AI that can observe its surroundings, understand instructions and actually move its hands. Madison Huang, senior director for Omniverse and robotics at NVIDIA, visited the site. Four days earlier, LG Group chairman Koo Kwang-mo and Jensen Huang had signed a memorandum in Santa Clara to widen their collaboration on physical AI, AI infrastructure and mobility.
Schaeffler — humanoid gearboxes move from machining to forming, volume production in 2027
Schaeffler, the German maker of bearings and other mechanical components whose main market is automotive, has completed large-scale validation of the strain wave gearboxes used in humanoid joints. It says it is ready to enter volume production in 2027. A strain wave gearbox delivers a high reduction ratio and high stiffness in a small space, and it sits at the centre of the actuator that moves a joint. That actuator accounts for roughly half the manufacturing cost of a humanoid. Making it by machining metal took time and capital investment. Schaeffler saw that method as a potential major bottleneck in bringing humanoids to volume scale.
The new method presses the part into its final shape in seconds under high force. Torque capacity and efficiency match those of the machined part, while manufacturing cost falls by more than 25% and material use by more than 75%. Process steps that took minutes take seconds. Schaeffler has supplied more than two million gearboxes made this way to automotive customers over the past decade. Production will start in Germany and expand to other regions. “We deploy humanoids along our entire global value chain and therefore have first-hand knowledge of the requirements of this technology and can develop key components based on our know-how,” says David Kehr, president of humanoid robotics at Schaeffler.
