In this article, I investigate the use of conventional single-level shuttle systems for goods-to-person picking of grocery items in e-com. I will explain why this particular type of automated storage and retrieval systems poses certain problems for eGrocery picking, and I will compare them to alternative concepts like AutoStore. It will become clear why the typical structure of eGrocery orders clashes with two properties of the fundamental design of conventional single-level shuttle systems.
Single-level shuttles are powerful automated storage and retrieval systems (AS/RS). Currently, there are no other systems that could deliver the same or higher storage and retrieval performance per square meter footprint for good-to-person picking systems. Ever since their inception they have taken the market by storm, replacing conventional Miniload stacker cranes and enabling material handling equipment (MHE) solution providers to build ever more dynamic systems. Some systems these days have as many as 3.000 individual single-level shuttles, and larger systems will surely come. Almost all major MHE companies have single-level shuttles in their portfolio, some even have multiple versions of them (cf. Table 1).
|SSI Schäfer||Cuby, FlexiShuttle|
|Knapp||OSR, OSR Evo|
As versatile high-performance AS/RS, single-level shuttles are great for many applications in many industries, most notably to power goods-to-person (GtP) picking systems. With the rise of eGrocery to the awareness of a larger consumer base, and with incumbent food retailers (finally) acknowledging that eGrocery is here to stay and won’t simply go away, it becomes increasingly relevant to understand if single-level shuttles are equally well-suited to power GtP systems in eGrocery as they are in other industries like general retail. As so often, however, things aren’t quite as simple in eGrocery.
So, how well do single-level shuttles perform in goods-to-person picking applications for eGrocery? Let’s look into some of the challenges we face when designing shuttle-based solutions for eGrocery.
eGrocery Order Structure
By tendency, eGrocery has a rather unique order structure. Think about how you order at Amazon or Zalando: you order one or two different products, and normally only one unit of each type of product you order. In logistics speech, you order one or two orderlines per order and one piece per orderline. Compare this to your weekly or bi-weekly grocery shopping (whether in-store or online): when buying groceries, you will often have between 20 and 50 different products (stock keeping units, SKUs). Of many products you buy only one unit, of some products you will buy two or three. So, the number of orderlines per order is significantly higher and for some products the number of units per orderline is likely to be higher, too.
Now, why is that important? Because it has implications on system design and performance, as you will see.
System Design: Three Approaches with Weaknesses
When you design a goods-to-person picking system with an automated storage and retrieval system based on single-level shuttles, the typical order structure of eGrocery systems can pose a challenge.
Generally, any goods-to-person picking station in the system should be able to work on any order that you release into the system for picking. This implies that each picking stations needs access to all SKUs (at least to all SKUs which are part of the GtP picking system; some SKUs might be kept outside that system for manual picking, for instance).
Let’s assume that we are dealing with a system with significant dynamic performance requirements; single-level shuttles are high-performance systems, hence they are, in fact, often used when large amounts of storage and retrieval operations need to be performed. For the sake of simplicity, let’s assume that capacity for 6.000 double-cycles per hour shall be available and that we intend to run the system at capacity. A typical single-level shuttle system with two lifts per aisle will be able to do roughly 750 double-cycles per hour, which means we will need eight aisles to meet the performance requirements. If we (optimistically) assume an average of 600 picks/h at each pick station with the typical 1.2 items/orderline, each pick station will process 500 storage totes per hour and we will have 6.000/500 = 12 pick stations in the system.
Variant 1: All Pick Stations are Served by All Shuttle Aisles
Let us look at the first variant. Here, all pick stations will receive source totes from all shuttle aisles. (This could be considered the default configuration for most GtP systems, at least outside eGrocery). And at first sight, this does not sound like a problem, does it? After all, all picking stations and all shuttle aisles will be connected through conveyors which transport storage totes between picking stations and shuttle aisles. Figure 1 illustrates the concept.
Now, here is the problem: If all 12 pick stations need to get storage totes supplied from all 8 shuttle aisles over a conveyor loops with evenly spread load across all aisles, this is not going to work. You cannot run 6.000 totes/h over a conventional conveyor loop used to connect shuttle aisles with picking stations, simply because one single loop’s effective dynamic capacity limit will be between 1.500 (conventional loop) and 2.500 totes/h (high-speed loop). Faster loops (up to 5.000 totes/h) do exist, yet the sensitivity of (a small) part of the grocery assortment would exclude such items from transportation at the highest pace. Hence, two to three high-speed loops (e.g., on different vertical levels) would be needed, making the concept financially unattractive and a mess to service.
Let’s remember the difference between eGrocery and other e-commerce business in terms of order structure. At a typical (general, fashion…) retailer you order one or two products. With this order structure, in a goods-to-person picking system you could complete any order at ease at any pick station because on average there is only one, sometimes two storage totes needed for any order. A concept as the one described above is very likely to work for general retail or fashion. With eGrocery, however, and an average of, say, 25 different items included in any one order, chances are that storage totes from ALL shuttle aisles would need to travel to one particular picking station for every order. The resulting traffic would exceed capacity of the conveyor loop very rapidly.
Variant 2: Division into Clusters
One alternative solution is that you divide the shuttle system up in clusters. A cluster could consist, for instance, of two to four aisles. Each cluster would hold the entire inventory needed at any picking station. With three picking stations per cluster, each station pulling 500 storage totes/h from the three aisles of the cluster, the conveyor could keep up with the performance requirements. Each cluster would be operated almost independently and inter-segment traffic would only occur if one cluster ran out of stock for a particular SKU. Figure 2 illustrates the concept.
A beautiful idea, albeit with one significant drawback: you have to replicate the entire inventory for each cluster. And a three-aisle shuttle system will not be able to hold a large range of SKUs; even with long and high aisles more than 10.000 SKUs per cluster would not be economical (you may be able to squeeze in some more, but then your increasing replenishment effort would begin to offset economic gains from GtP picking). And if we are honest: replicating inventory multiple times does not sound like a great idea, let alone for food products with best-before date.
Variant 3: Forwarding Order Totes
What else could we do? We could maintain efficient inventory levels and scrap the cluster idea while forwarding order totes from pick station to pick station. Each pick station would pick the SKUs that can be conveniently delivered by the closest shuttle aisle and then for SKUs from more remote shuttle aisles forward the order tote to a closer pick station. This approach avoids the disadvantages from the concepts discussed before. Figure 3 illustrates the concept (please note that order totes are not included in Figures 1 and 2 for the sake of simplicity, but are included in Figure 3 for illustration).
In this scenario, however, synchronization between storage totes and order totes at pick stations becomes really difficult. Significant (sequencing) buffers will be needed in the conveyor layout for both, order totes roaming between pick stations as well as storage totes coming from shuttle storage. For systems with high dynamic performance requirements, this does not seem to be a feasible approach.
Sequencing Requirements vs. Performance, and the Problem of Uneven Workload
AS/RS systems generally face a trade-off between performance and sequencing requirements. This is true for all types of AS/RS, yet it is a larger problem for single-level shuttles than it is for conventional stacker cranes or hive robot systems (like AutoStore or Ocado). In the moment you want to retrieve storage totes in a particular sequence, the system is highly likely to deliver fewer units per hour than if you let it decide itself which sequence is best.
So far, so good. But why is this a bigger problem for single-level shuttles than for other systems? It is because each shuttle can only access SKUs in its particular storage level (remember: not all SKUs are available in all aisles, let alone in all storage levels). If you let the „shuttle controller“ decide which SKUs to retrieve in which order, the controller could place emphasis on equal utilization of all shuttles. If, however, you enforce a retrieval sequence from a shuttle system, some shuttles could be overloaded with retrieval jobs, unable to keep up, while other shuttles will idle, reducing the effective utilization of the system’s capacity. Sales engineers with a sense of reality therefore plan their systems with average capacity utilization of the shuttle as low as 40%.
This is a general challenge with shuttle systems, i.e. in all industry segments. In eGrocery, however, we do have sequencing requirements which we cannot ignore. They result from the fact that different products are subject to different sensitivity and to different weight. Product weight and product sensitivity result in a number of so called crash classes, dictating which products you need to pick first (because they are heavy and relatively stable), which products you pick last (because they light and sensitive to pressure, i.e., they can easily break) and everything in between. The sequence requirements are not absolutely strict – you can mix products from within the same crash class. But even so it is one additional sequencing requirement that comes on top of the sequencing requirement resulting from the particular order structure: you want to „these 25 different products“ to go into this order tote more or less in one go; you don’t want to only one or two items belonging to one order now, some more items in 30 minutes and some later in the afternoon.
The problem can be alleviated or exacerbated by choice of pick station design. Most goods-to-person pick stations have one pick-from location and two, four or even more pick-to locations. The advantage of such designs are twofold: Firstly, you can make use of the so-called batch factor: if the same product is needed by more than one order, one (or: fewer) retrievals of a storage tote from the AS/RS will suffice. Secondly, since more than one order tote is present in the pick station, the retrieval sequence from the AS/RS can be less strict: storage totes for any of the orders present in the pick station can be retrieved without breaking the required sequence for any of the orders. If, however, the user of the automated system insists on 1:1 pick stations (as many eGrocery retailers, in fact, do in their search of the highest possible pick rate), sequencing requirements will become so intense that there needs to be massive overcapacity in the shuttle system to prevent pick stations from starving (i.e., from idle time due to waiting for storage totes to pick from).
This, by the way, is a problem that hive-robot systems don’t have to the same extent: while they, too, are subject to the trade-off between performance and sequencing requirements, the bots can always be utilized at capacity since they roam the grid and are not confined to one storage level like a single-level shuttle. They will become less efficient since they will spend more time digging out totes when a particular sequence is enforced, but the difference will be much smaller than for a shuttle system.
How about Shuttle-based Microfulfillment Centers (MFCs)?
Shuttles are the technological base for some of the so-called Microfulfillment Centers (MFCs) as offered, for instance, by companies like TakeOff (in cooperation with Knapp) and Dematic. MFC is nothing but a fancier name for small-scale automated solutions for more decentralized fulfillment of (mostly, but not exlusively) online grocery orders. While until recently, logistics engineers and sales managers with a sense of responsibility told their prospective customers that shuttle systems will need at least (say) 10m of clear height in order to be able to deliver the performance needed for a reasonable business case, it seemingly has become fashionable to try to sell these miniature shuttle-based goods-to-person picking solutions for installation in backrooms of retail stores. There is much to be said about this concept, and a separate article is needed for a comprehensive analysis of the concept and its economic viability. Only one aspect shall be of interest for the present discussion: their small size exacerbates the problem of even capacity utilization of shuttles, as discussed above.
The shuttle-based MFC design currently so extensively marketed is intended for buildings with 5m of clear height, allowing 10 shuttle levels per aisle. The archetypical system will include three aisles, two of which will hold ambient temperature range products, the third one chilled temperature range products. Each shuttle aisle would contain short of 3.000 storage locations (i.e., 6.000 storage locations for ambient products and 3.000 for chilled products). This very low number of locations does not allow for large amounts of inventory per SKU, hence it must be assumed that for each SKU a maximum of one storage tote of inventory will be available (one MFC provider has gone so far as to suggest that 9.000 storage locations will be enough for more than 12.000 SKUs due to the use of compartmentalized storage bins. I am rather certain, however, that the stock management and replenishment pain resulting from the use of compartmentalized bins will outweigh any economic benefit from having more SKUs in the automated system).
One peculiar aspect of MFC design is that fast-moving products are kept in the automated system while the long-tail will be picked from the shelves of the conventional store. Under normal circumstances, however, the normal (and arguably way more efficient) way would be to move the long tail into the automated system and pick fast-movers manually. The fact that in an MFC, the fast-moving products will be placed in the automated system has implications. Fast-moving products – the early, very steep part of the Pareto curve – are subject to significantly different demand. It is an hypothesis which requires backup via simulation of observation of running systems, but it seems reasonable to assume that load balancing of shuttles within one aisle, as well as between aisles, will become even more difficult, further reducing effective capacity utilization. With only 10 storage levels (and thus 10 shuttles) per aisle, the result would be that the pick station will frequently starve, waiting for new storage totes to arrive.
Shuttle systems are powerful AS/RS, and they are popular for a reason. The high performance does not come for free, though: shuttles and racking for shuttles are expensive. All the more it is important that shuttle systems are planned properly so that they can deliver a good business case for the user. The design of shuttle-based goods-to-person picking solutions does not seem to be well understood, however, and even much less so in the generally little-understood segment of eGrocery. Single-level shuttles can be the backbone of great eGrocery solutions, but while looking great on paper, their effective performance can easily disappoint if it is not fully understood which parameters influence their performance.
The application of single-level shuttles to goods-to-person picking in eGrocery applications is much more difficult than it seems at the outset. Firstly, the very peculiar order structure of online grocery orders plays an important role for the performance of shuttle systems. It clashes with the segmentation of storage and retrieval capacity in (1) aisles and (2) levels, which is inherent to the fundamental design of single-level shuttle solutions. Secondly, the discussion of the three design approaches has illustrated that even where the low effective capacity utilization of the shuttle system can be compensated for by the installation of additional aisles and shuttles, the resulting traffic of storage and/or order bins on the front zone conveyors connecting the goods-to-person picking station will represent an obstacle to operations and order release planning, with the only alternative being a replication of inventory, which seems undesirable in general and for groceries in particular.
While many people in the MHE industry were initially puzzled that Ocado, the British online supermarket and automation solution provider, would base their automated solution entirely on a cube storage powered by hive robots, their solution has two big advantages over single-level shuttle systems: performance of each hive robot can be fully utilized at all times, and therefore system performance is rather predictable, both of which is not true for shuttle systems. And what applies to Ocado applies to AutoStore, too. There is a reason that AutoStore and their distributors, most of whom until shortly before the pandemic did not even spend any effort on marketing their solution to eGrocery companies, have closed a number of deals with food retailers (like US-based HEB) for Microfulfillment Solutions. And the smaller the system, the more likely AutoStore will have an edge over a shuttle-based solution due to easier control and better effective capacity utilization. So, it seems fair to say that AutoStore is the better-suited AS/RS for Microfulfillment Centers. Single-level shuttle-based systems are likely to outperform AutoStore in larger systems when they can operate at their sweet spot with the right combination of throughput and SKU count, but then, again, resulting front zone traffic will be difficult to manage. For small systems, however, the strict segmentation of storage and retrieval capacity in single-level shuttle systems represents a distinct disadvantage in comparison to cube storage systems with hive robots, as well as to multi-level shuttle systems, the latter being an alternative that somehow seems to have been forgotten, or at least neglected, by many providers of automated systems.
This article intends to support both, suppliers and customers of shuttle-based GtP picking solutions, by discussing important design aspects of such solutions for eGrocery applications. Do you have feedback to this text, or do you want to discuss solution design for online grocery fulfillment? Reach out to firstname.lastname@example.org.