In short: Travel is the largest single component of manual picking time, and the sequence most systems hand a picker is a location code sort with a serpentine flag, which is the traversal rule wearing a different name. Traversal holds up at high pick density and loses badly below roughly one pick per aisle, where largest gap and composite rules do better. Ratliff and Rosenthal showed in 1983 that the exact optimum is computable for a single block warehouse, so the heuristic is a choice rather than a necessity. Batching normally returns more than routing because it changes the number of trips instead of their shape, and the bill for it lands downstream in sortation and packing. Measure walked path against optimal path for the same location set, which separates the routing question from the slotting one.
The task list is sorted ascending by location code, so the picker enters aisle 3 at the near end, walks its full length for two cartons, leaves at the far end, crosses to aisle 7 and does the same again. Nobody in the building chose that pattern. It is what happens when the pick sequence is a sort on a location string, and the string was designed to be readable rather than walkable.
The breakdown reproduced in Tompkins and colleagues' Facilities Planning (2010) puts travel at roughly half of an order picker's time, with searching, extracting and administration splitting the rest. De Koster, Le-Duc and Roodbergen's 2007 review in the European Journal of Operational Research reaches the same place from the design side: travel is the component worth attacking because it adds nothing and it is the part a planner controls.
Put ordinary numbers on it. Forty pickers at seven and a half productive hours is 300 hours a day, of which about 150 are walking. Suppose a change of routing rule removes a tenth of that walk. Fifteen hours come back each day, which is two heads, and around 3,750 hours across a working year. The saving is available without buying anything, which is why it is worth understanding what your system is doing before you go looking for a module.
The route your system gives you is a sorted location code
Ask a WMS vendor whether the product optimises the pick path and the answer is usually yes. Ask what the rule is and the answer is usually a sort by aisle, then bay, then level, with a flag that reverses the direction on alternate aisles. That reversal is the whole of the optimisation.
The resulting pattern is the traversal rule, also called the S-shape: enter every aisle containing at least one pick, walk it end to end, move to the next one. It has two properties that keep it popular. It is trivial to compute, and it produces a route a person can follow without thinking, which matters more than planners tend to credit.
It also has one failure mode, and it is severe. A picker who enters an aisle for a single carton near the entrance still walks the full length of that aisle, because the rule commits to a complete traverse. In a building with 40 metre aisles and a task that touches nine aisles for eleven lines, the traverse discipline is spending most of the walk on aisles where there was nothing left to collect.
Five routing rules and where each one earns its place
The literature settled the comparison a long time ago. Hall's 1993 paper in IIE Transactions gave distance approximations for the main heuristics, and Petersen's 1997 study in the International Journal of Operations and Production Management simulated them against each other across order profiles. The rules worth knowing:
Traversal, or S-shape. Full traverse of every aisle with a pick. Simple, predictable, and efficient when most aisles entered contain several picks.
Return. Enter each aisle from the front, walk to the deepest pick, come back the way you came. Wasteful for deep picks, strong when picks cluster near the cross aisle.
Midpoint. Aisles are split in half. Picks in the front half are collected from the front cross aisle, picks in the back half from the back one. It beats traversal at low pick density and is easy to explain on the floor.
Largest gap. The picker enters an aisle, collects up to the largest unvisited stretch, and reverses out, so the longest gap in each aisle is never walked. It is the strongest of the simple heuristics at low density and the hardest to follow without system guidance.
Composite. Decides aisle by aisle whether to traverse or return, based on where the next pick sits. It needs a small amount of lookahead and it removes most of the gap between the simple rules and the optimum.
Above these sits the exact answer. Ratliff and Rosenthal showed in Operations Research in 1983 that order picking in a rectangular single block warehouse is a solvable case of the travelling salesman problem, computable by dynamic programming in time linear in the number of aisles. Roodbergen and de Koster extended the treatment to multiple cross aisles in the International Journal of Production Research in 2001. The exact route is available on ordinary hardware for realistic pick lists, and it typically sits a few percent below the best heuristic rather than transforming the number.
Pick density decides which rule you should be running
The single parameter that sorts the heuristics is pick density, meaning picks per aisle visited. Compute it from a month of task history: total lines in a task, divided by distinct aisles touched in that task, averaged over tasks.
Above roughly four picks per aisle, traversal is close to optimal and the alternatives buy little, because you were going to walk most of the aisle anyway. Below about one and a half, traversal is walking long stretches of empty rack, and largest gap or composite pull ahead by a margin worth having. Petersen's simulation work puts the spread between the best and worst policy at a substantial fraction of travel distance in the low density region, which is exactly the region most e-commerce operations live in.
Here is the part that gets missed. Pick density varies by zone, by hour and by order type in the same building. Case picking to store orders might run at six picks per aisle while the single line web orders picked from the same reserve area run at 0.4. One routing rule applied across all of it is optimal for neither, and the fix is a rule chosen per task type rather than a global setting.
Work out your own crossover before changing anything. Take a week of pick tasks, group them into density bands, and estimate route length under each rule for the tasks in each band. The comparison is arithmetic on data you already hold, and it tells you whether the rule you are running is wrong and where.
Batching is usually worth more than routing
Routing changes the shape of a trip, while batching changes how many trips there are, and the number of trips is normally the larger lever.
A single line order picked on its own means one trip for one line. Ten such orders picked together as one task with a tote per order means one trip for ten lines. The travel per line falls by something close to an order of magnitude, and no routing rule can compete with that because the trip itself is the cost.
The problem is harder than it looks. Gademann and van de Velde showed in IIE Transactions in 2005 that order batching to minimise total travel is NP-hard once batch size exceeds two, so practice runs on heuristics. Two families dominate. Seed algorithms pick a starting order and add the orders that share the most aisles with it, described by Elsayed and Unal in the International Journal of Production Research in 1989. Savings algorithms adapt the Clarke and Wright construction from Operations Research in 1964, computing the travel saved by combining any two orders and merging greedily down the list.
Both are cheap to run and both beat first-come-first-served grouping, which is what most sites do by default when they let the release sequence define the batch.
The cost that makes batching a decision rather than a free win sits downstream. Picking ten orders into one container means somebody sorts them afterwards, at a put wall, a sorter, or a packing bench with a stack of totes. Sortation labour rises roughly with lines batched, while travel savings flatten out as batch size grows. There is a batch size where the two curves cross, it is specific to your pack area design, and it is usually smaller than the number the picking team would choose if left alone.
Congestion is the cost the route model cannot see
Every routing model above assumes an empty building. Real aisles contain other pickers, replenishment trucks and a pallet somebody left mid aisle.
Gue, Meller and Skufca studied this directly in IIE Transactions in 2006, looking at pick density in narrow aisle areas, and found that congestion effects become material well before an area looks busy, because blocking is a function of how often two pickers want the same aisle segment rather than of average occupancy. Their result matters for planning because the standard response to a throughput problem is to add pickers, and in a congested area the marginal picker adds less than the average picker did.
Two practical consequences. A routing rule that concentrates pickers in the same fast moving aisles will hit congestion earlier than one that spreads them, so the optimal route on paper can be the slower route on the floor. And any measurement of a routing change made during a peak week will be contaminated by congestion, which is a good reason to test rules in a quieter period first.
The measurement that separates routing from everything else
Lines per hour will not answer this question, because it moves with order profile, staffing mix and replenishment reliability. Travel distance per line is a better instrument and it belongs to the slotting conversation, which X5 covers.
For routing specifically, the number to build is a ratio. Reconstruct the sequence the picker actually walked from WMS task records, which carry location and timestamp in sequence. Derive coordinates for each location by parsing the aisle, bay and level out of the location string against a rack drawing. Compute the length of the walked path, then compute the optimal tour over the same set of locations for that task. Divide one by the other.
That ratio isolates the routing decision. A task whose walked path is 1.6 times the optimal for the same locations was routed badly, whatever the slotting looks like, and no amount of moving items around the pick face will recover that particular loss. Plot the distribution rather than the mean, group it by task type and by zone, and the tasks in the tail will point at a specific rule applied to a density band it does not suit.
Where this stops
The honest limit is that pickers deviate. Elbert, Franzke, Glock and Grosse looked at exactly this in Computers and Industrial Engineering in 2017, modelling what happens to routing policies when pickers depart from the given route, and the finding is uncomfortable for anyone selling optimal paths. Deviation erodes the advantage of the more sophisticated policies, and the simple traversal rule turns out to be relatively resistant to it because a picker following it is rarely tempted to improvise.
That gives a real design constraint. A largest gap route computed to the metre is worth nothing if the picker cannot see why it doubles back, decides it is wrong, and walks their own path. The rule you can actually run depends on how the route is presented: a voice or RF system that gives one location at a time gets compliance, a printed list invites editing.
Two further boundaries. Routing optimisation walks a given layout efficiently, so a pick face that reflects an obsolete range will still be walked further than it should be, and that is a slotting problem before it is a routing one. And the exact algorithms assume a rectangular arrangement of parallel aisles with cross aisles at the ends; a building with mezzanines, angled aisles or a pick module with lifts needs the graph modelled properly rather than an aisle count fed into a formula.
Pull one week of pick task records, derive coordinates from the location codes, and compute the walked path against the optimal tour for the same locations on every task; the shape of that ratio will tell you within a day whether your routing rule or your slotting is the thing costing you the walk.