IN Brief:
- Six Sereact robotic picking systems will be installed at Arvato facilities in Dortmund, Gütersloh, and Memphis.
- The systems will pick from AutoStore bins into cartons and totes, and transfer items between conveyor lines.
- Operating data from three sites will test whether adaptive robotic picking can support a repeatable network standard.
Sereact will deploy six AI-guided picking robots across three Arvato fulfilment centres in Germany and the United States, extending an initial implementation into a multi-site operating programme.
Four systems will be installed in Dortmund, with one unit assigned to Gütersloh and another to Memphis, Tennessee. Two of the sites will use the robots to transfer products directly from AutoStore bins into cartons or totes, while the third will handle items moving between conveyor lines.
The installations will run Sereact’s Cortex 2.5 software, which combines machine vision and robotic manipulation to recognise unfamiliar items, select a suitable grasp, and adjust when packaging, presentation, or product mix changes. Rather than relying on a fixed catalogue of pre-programmed stock-keeping units, the system is intended to handle a changing assortment through a common software layer.
Dortmund will carry the largest concentration of equipment and provide the main high-throughput deployment. Gütersloh and Memphis will broaden the test across different facilities, order profiles, and integration environments, giving Arvato a basis for comparing performance beyond a single warehouse.
“Integrating Sereact’s AI-powered robots directly at our AutoStore ports and conveyor lines allows us to automate picking with a level of flexibility that conventional systems simply can’t match,” said Markus Billmann, senior expert for process automation at Arvato. “The technology adapts to our changing product mix in real time, and scaling it across multiple sites gives us a consistent automation standard across our network.”
Sereact has been expanding its commercial operations since completing a $110 million Series B funding round in April. Its technology is also being used in automotive and industrial settings, where customers include Daimler Truck, Mercedes-Benz, and BMW.
The final manual transfer
Automated storage has removed much of the walking and searching from modern fulfilment, yet many systems still present a tote to a person for the final item transfer. The storage grid can sequence inventory rapidly, but the last movement into a carton, tote, or conveyor lane remains dependent on visual identification, dexterity, and exception handling.
That transfer is difficult to automate across a broad assortment because products may be soft, reflective, transparent, deformable, loosely packed, partially obscured, or entangled with neighbouring items. A gripper that handles rigid cartons successfully may struggle with textiles, bags, cylindrical containers, or products that shift as the robot approaches.
Adaptive picking software is intended to reduce the amount of item-specific engineering required. Instead of teaching the robot every product in advance, the system interprets the scene, chooses a grasp, and modifies its action when the first attempt is unsuitable.
The commercial measure is not whether the robot can pick an individual demonstration item, but whether it can sustain useful throughput across live order waves while keeping failed picks, product damage, and manual interventions within acceptable limits. Exceptions must also be removed without blocking the AutoStore port or interrupting the surrounding conveyor flow.
Arvato’s use of the same platform in two operating configurations will expose the software to distinct conditions. AutoStore presents inventory in a controlled bin at a defined workstation, whereas conveyor handling introduces movement, spacing, orientation, and timing variables that can change from one item to the next.
Multi-site standardisation can simplify training, software support, spare-parts planning, and performance comparison, but each installation still has to connect with local warehouse-management logic, carton selection, order sequencing, labelling, quality checks, and packing processes. A standard robotic cell cannot compensate for inconsistent master data or poorly designed exception routes.
Another adaptive system has already moved beyond the laboratory, with Geek+ extending robotic-arm picking into mixed warehouse operations. Both deployments show how suppliers are attempting to replace highly constrained robotic applications with systems that tolerate a wider range of everyday variation.
Labour availability provides a strong commercial incentive, particularly where repetitive picking roles are difficult to recruit and retain. The calculation nevertheless extends beyond hourly labour cost, because the robot’s utilisation, availability, maintenance, integration, and supervision determine how much productive work it absorbs over a full shift.
Upstream and downstream balance will be equally important. A fast picking arm offers little benefit when bins arrive irregularly or packing capacity cannot accept its output, while a prolonged robot stoppage can strand inventory at an automated storage port that was designed to remain productive.
Data from the three sites should allow Arvato to separate local operating issues from platform-wide performance. Pick success by product type, intervention frequency, damage, recovery time, and sustained rate will show whether the technology behaves consistently when the assortment and workflow change.
The measured scale of the rollout is useful: six systems are sufficient to expose repeatability problems without committing the wider network before the operating evidence is available. If the common software model holds across Germany and the United States, Arvato will have a clearer route from isolated robotic cells towards a repeatable fulfilment standard.
