Retrofitting GTP Stations with Robotics – Answers from the even Webinar with Berkshire Grey

In a nutshell: Anyone who has a Goods-to-Person station and wants to automate picking faces a question that is more complex than it sounds: What can actually be picked – and what cannot? Nicolas Dewit from Berkshire Grey explained in the even Webinar how the decision is made, what really matters during preparation, and why the business case is established within one to two years in most cases. And why you shouldn't necessarily plan for 650 picks per hour.

Let's imagine I have a GTP station. Today, a person is standing there. I want to free them up for other tasks and place a robot there instead. Is that straightforward?

Nicolas: That would be nice, but unfortunately not. Every GTP station offers fundamentally good conditions for robotics – but the robotic system alone doesn't make the difference. The whole setup needs to be defined. And you have to keep in mind: not everything can be automated. Some items will always need to be picked manually. The big challenge, therefore, is not the technology – it's the concept.

What do you look for when you assess a GTP station? What makes it a good candidate?

Nicolas: First of all, you need space. Many customers are tight on space, and then it becomes difficult to position the robot at all. The robot also needs to be fixed in place. Then the bin compartmentalization matters: the compartments within the bin must not be too close together, as this makes it harder for the robot to grip individual items. We have tested this with AutoStore up to 16 cells per bin – and it works. And of course, the distance between the source and destination bin is relevant. At Berkshire Grey, we can bridge up to 1.8 meters – beyond that, it's not feasible.

What is pickable – and how do you determine that?

Nicolas: Pickability is not binary. An item may look simple on paper, but the packaging doesn't suit it – for example, a plastic film that doesn't allow for a proper seal when gripped. Or items that deform significantly during picking. Or items that are so jumbled together in a bin that the robot cannot safely extract a single item. This then leads to double or even triple picks.

How do you find that out? Are length, width, height, and weight from the master data sufficient?

Nicolas: That's the first step. We always start with an approach based on master data and look at the size and weight of the items and how the packaging is described. Based on this, we can already make an initial segmentation and identify what we would rule out straight away. But that's not enough. The second step is a real test under actual conditions – ideally on-site at the customer's facility. We bring in a machine, run tests for a few days, and then determine what actually works and what remains an exception.

If I find that some items cannot be automated – does that kill the business case?

Nicolas: No, an exception must not prevent the project from moving forward. What's important is that the exception handling concept is defined from the outset. What we typically do: non-pickable items are automatically routed via conveyor technology to a manual picking station. Or the robot tries up to three times – if it doesn't succeed, the item goes to the manual station. What exactly is done is discussed on-site with the customer. But the concept must be in the plan before the first robot is installed.

What does the customer need to prepare structurally and technically?

Nicolas: The customer generally doesn't need to worry too much about space – we take care of everything. The robot needs to be fenced off, it requires a dedicated power supply, and the footprint is manageable: we're talking about three to five square meters including the cell. As for the software integration: we communicate via API with the customer's WMS or WCS. We have not yet encountered a case where we had to say we couldn't integrate with a system.

How many picks per hour are realistic?

Nicolas: That's a tricky question. Many customers fixate from the start on figures like 650 or 800 picks per hour and want them guaranteed. Yet there are external factors that influence this – for example, how quickly the AutoStore system triggers the next bin. If the robot picks fast but has to wait for the next bin, gripper speed alone doesn't help much. Our range is between 600 and 800 picks per hour – but this needs to be tested. With AI, the systems also improve over time: they grip better and place items more accurately. The placement precision in the destination bin also makes a significant difference.

Are there weight limits?

Nicolas: We are at 6 kilograms per item. That is actually already quite substantial – if a bin holds 30 kilograms, at 6 kilograms per item you can still fit five pieces in.

When does such a retrofit pay off?

Nicolas: Typically between one and two years, depending on the business case. We offer this as both a CapEx and an OpEx model – including as RaaS. The main drivers are rising labor costs, workforce shortages, and the opportunity to get more out of an existing GTP investment. What's important: we won't necessarily try to automate 100 percent of all picks. We ensure that we cover the economically relevant volume – and on that basis, the project pays off.

What should someone do first if they want to explore this topic now?

Nicolas: First, define the process and the objective: what do I want to achieve, what needs to be automated, how do I measure success? Then analyze the SKUs and volume data: what are my hot items, what do I sell the most? After that, get in touch with us, test the items – and then we work through the business case together, look at labor costs and other factors, and decide whether and how it fits.

Conclusion

What the conversation with Nicolas shows: the step from manual to robot-assisted picking at a GTP station is achievable – but it starts with the right concept, not with the purchasing decision. Anyone who clarifies upfront which items are pickable, how exceptions are handled, and what the material flow looks like is laying the foundation for a project that pays off. Conversely, those who start with unrealistic throughput expectations or only think about the exception handling concept after installation will run into problems – as experience consistently shows.

The recording of the Webinar is available here.