AI Route Optimization in the Warehouse

Walk the floor of any manual or semi-automated distribution center during a busy shift and you will see the single largest hidden cost in the whole operation moving right in front of you: people walking. Not picking, not scanning, not packing. Walking. Study after study, and my own stopwatch on more than one warehouse floor, puts the travel component of an order picker's day somewhere between forty and sixty percent of paid time. That is time the customer never sees and never values, and it is the first thing a serious optimization effort should attack. This guide is about the technique that attacks it most directly: AI route optimization, the practice of sequencing the stops on a pick list so that a picker visits every required location in the shortest total travel rather than in whatever accidental order the list happened to print. It sits inside the wider automation picture covered in the warehouse automation complete guide, and it is one of the few automation wins that needs no new conveyor, no robot and no capital project to deliver, only better software making a better decision.

The message up front: route optimization is the cheapest large win in warehouse operations because it changes the sequence of work, not the physical plant. You are not buying steel or robots, you are asking the software you already run to make a smarter decision about the order in which a picker walks. On a large-aisle facility that can cut travel by a quarter to a half, and travel is where most of the labour hides. For where this fits in the wider automation stack, start with the warehouse automation pillar.

1. The travel problem in picking

Order picking is the most labour-intensive activity in most warehouses, and travel is the most labour-intensive part of picking. When a picker receives a list of ten lines, each line lives at a physical location somewhere in the racking. The picker has to physically go to each of those ten locations, retrieve the quantity required, and move on. The picking itself, the reaching and grabbing and scanning, takes a few seconds per line. The travel between locations takes far longer, and it scales with the size and layout of the building. In a compact shelving room the travel might be trivial. In a large distribution center with long aisles and cross-aisles, a single order can involve hundreds of metres of walking, and a picker on a busy day walks the equivalent of several kilometres.

Here is the part that makes route optimization worthwhile rather than academic: the amount of walking required to satisfy a given pick list is not fixed. It depends entirely on the order in which the picker visits the locations. Visit them in the sequence they happen to appear on the printed list, which is usually the order the lines were keyed or the order of the item numbers, and you can easily walk twice as far as necessary, doubling back down aisles you already passed, criss-crossing the floor because location A0107 and location A0109 printed pages apart even though they sit two bays from each other. Visit the same set of locations in a sequence that respects the geometry of the building, and you walk the minimum. Same items, same quantities, same picker, dramatically different travel. That gap between the naive sequence and the optimal sequence is the entire prize.

Because travel is such a large slice of picking labour, and because picking labour is such a large slice of total warehouse cost, even a modest percentage reduction in travel translates into real money and real throughput. Cut travel by a third on an operation where travel is half the shift and you have freed roughly a sixth of your total picking labour without hiring, without capital, and without touching the physical building. That is the arithmetic that makes routing the first optimization most operations should reach for, ahead of the expensive automation projects that dominate the trade press.

2. How route optimization works

At its heart, route optimization is a sequencing problem. You have a set of locations that must all be visited, a fixed start point (usually the pick station or the mouth of the aisle block) and a fixed end point (usually the drop-off or packing area), and a known travel cost between every pair of locations determined by the warehouse layout. The task is to find the order of visits that minimises total travel distance while visiting every required location exactly once. Anyone with an operations research background will recognise this immediately: it is a constrained version of the classic travelling salesman problem, adapted to the grid geometry of aisles and cross-aisles rather than an open plane.

The difference between a naive walk and an optimized route is easiest to see than to describe. The diagram below shows the same eight pick locations served two ways. On the left, a naive sequential walk that visits locations in list order, doubling back and crossing aisles. On the right, an optimized path that groups the stops by aisle and threads them in a single efficient sweep, ending near the drop-off.

The same locations, the same items, the same picker. The left path zig-zags between aisles and doubles back on itself because it follows list order. The right path groups the stops by aisle, walks each aisle once in a single up-or-down pass, and finishes near the drop-off. The saved distance is the shaded difference in walking, and across thousands of orders a day that difference is a headcount. The engine that produces the right-hand path has to know the physical coordinates of every location, the graph of aisles and cross-aisles that connect them, and the travel cost of every possible move, then search the space of orderings for the cheapest one.

3. Pick routing methods

Not all routing is AI, and it is worth being honest about the spectrum. Warehouses have used routing heuristics for decades, long before anyone attached the word intelligence to them. These heuristics range from the trivially simple to the genuinely optimal, and each trades travel efficiency against computational and operational complexity. The table below lays out the main methods in the order they typically appear as an operation matures.

The practical lesson from this table is that you do not have to jump straight to the most complex method to capture most of the benefit. Moving an operation off raw sequential order and onto a disciplined S-shape heuristic often captures the majority of the achievable travel saving for almost none of the implementation effort. The optimal solver squeezes out the last portion, and on a large, high-volume operation that last portion is still worth chasing, but the honest sequencing is to grab the cheap heuristic win first and reserve the solver for where the volume justifies it.

4. From simple heuristics to AI optimization

So where does the AI actually come in, and where is it just a fashionable label on an old idea? The heuristics in the table, the S-shape and return methods, are fixed rules. They do not compute the true optimum, they apply a sensible pattern that is close enough most of the time and, crucially, easy for a picker to follow and for a supervisor to understand. The optimized method is different in kind: it treats each pick list as an instance of a shortest-path search and computes a route that is genuinely close to the mathematical minimum for that specific set of stops.

Because the underlying travelling salesman problem is computationally hard, solving it exactly for a large pick list in the split second a warehouse needs is not practical, so real systems use approximation algorithms and metaheuristics: nearest-neighbour construction, two-opt and or-opt local search, savings algorithms, and where the problem is large or richly constrained, techniques such as simulated annealing, genetic algorithms and increasingly machine-learning-guided solvers that learn good routing patterns from historical order data. These produce routes within a few percent of optimal in milliseconds, which is exactly the trade a live operation needs.

The genuinely new contribution of AI, as opposed to classical operations research, is twofold. First, learning from history: a model trained on months of order data can anticipate which items tend to appear together and pre-shape routes and even storage to suit, rather than treating each order as a blank slate. Second, handling the messy constraints that pure distance minimisation ignores, congestion between pickers, aisle capacity, weight and stackability of the pick, refrigeration zones, and the changing mix of orders through the day. A classical solver minimises distance; a well-built AI system balances distance against the operational realities that make the shortest-on-paper route not the fastest in practice. That is a real distinction, and it is also where a lot of vendor marketing overreaches, so keep a sceptical eye on which of the two you are actually being sold.

The honest limitation: a route that is optimal on distance is not automatically optimal in the building. It ignores the other picker coming the other way in a one-person aisle, the heavy item that should be picked last so it sits at the bottom of the tote, and the chilled line that has to be grabbed in one trip. A distance-only optimizer can produce a mathematically shorter route that is slower or unworkable on the floor. The value is in balancing distance against these constraints, not in worshipping the shortest number.

5. Combining routing with batching and slotting

Routing is powerful, but it is one of three levers that work on the same underlying travel cost, and treating it in isolation leaves most of the prize on the table. The other two are batching and slotting, and the biggest gains come from using all three together rather than optimising any one to death.

Batching groups several orders into a single pick trip so the picker collects for multiple orders in one walk of the aisles, then sorts them into separate orders at a pack station. Routing then optimizes the combined stop list. Batching multiplies the effect of routing, because the fixed travel of a walk down an aisle is now amortised across several orders instead of one. The trade-off is the sortation work at the end and the constraint that batched orders must be compatible in timing and destination. Batching and routing are natural partners: batching decides which orders share a trip, routing decides how that trip is walked.

Slotting is the placement of items in locations so that fast-moving and frequently-co-ordered items sit close to each other and close to the dispatch area, shortening the routes before any sequencing happens. Slotting attacks travel at the source: no routing algorithm can make a well-travelled route out of items scattered to the four corners of the building, but good slotting means even a mediocre route is short because the popular items are clustered. Routing optimizes the path given the layout; slotting optimizes the layout so the paths are short to begin with. The two reinforce each other, and the most effective operations tune slotting periodically as demand shifts, then let routing extract the daily optimum from that improved layout. I have written separately on AI slotting optimization and on the wider picture of AI for picking optimization, and the honest summary is that routing, batching and slotting are three views of the same travel-reduction problem and should be governed as one program, not three disconnected projects.

6. The honest limits: congestion, layout and real-world messiness

Every route optimization pitch shows a clean before-and-after where the optimized line is obviously shorter, and on paper it always is. The floor is less tidy, and a practitioner owes you the caveats. The first is congestion. A distance-optimal route computed independently for each picker assumes each picker has the aisles to themselves. Run twenty pickers on individually optimal routes in a building with narrow aisles and you can route several of them into the same bottleneck at the same moment, and the aisle that is shortest on the map becomes the slowest in practice because everyone is queuing in it. Real systems have to coordinate across pickers, spreading load across aisles, which sometimes means giving an individual a slightly longer route so the whole floor flows better. Optimising each picker in isolation can make the collective worse.

The second limit is layout. Routing works with the building you have. If the aisle structure is awkward, if cross-aisles are few and far between, if the dispatch area is badly placed relative to the fast-movers, there is a ceiling on how short any route can be, and no algorithm lifts it. Beyond a point the answer is not better routing but a physical re-layout or a slotting overhaul, which is a bigger and slower project. It is important to tell a client honestly when they have hit the routing ceiling and the next gain is structural, not algorithmic.

The third is the everyday messiness that clean models omit: a location that is short-picked so the picker has to divert to a reserve location, a blocked aisle from a spill or a maintenance job, a replenishment truck parked mid-route, an item that turns out to be damaged, a picker who knows a shortcut the map does not model. A route that is optimal in the abstract can be defeated by any of these, and a system that cannot re-plan on the fly when reality diverges is brittle. The mark of a good routing implementation is not the elegance of the optimal route it computes at the start of a trip, it is how gracefully it recovers when the trip does not go to plan. That graceful re-planning under disruption is the same challenge that shows up in AI workforce route and dispatch optimization across the wider operation, and it is where the honest engineering, rather than the demo, lives.

7. Routing in the WMS

Route optimization is not usually a standalone product. In practice it lives inside the warehouse management system, because the WMS is the only place that already holds the two things routing depends on: the live pick lists coming out of order management, and the location master that describes where everything physically sits. A routing engine bolted on separately, fed by nightly extracts, is always a step behind reality and cannot react to a short-pick or a re-slot until the next sync. Native routing inside the WMS sees the same data the pickers see, in real time, and can re-sequence a trip the moment something changes.

This is also why routing quality is downstream of data quality. The engine can only compute a good route if the location master is accurate: correct coordinates, correct aisle and bay assignments, a faithful model of the cross-aisle graph and the walking distances between points. Warehouses that never invested in an accurate spatial model of their own building cannot route well no matter how sophisticated the solver, because the solver is optimising over a map that does not match the floor. The unglamorous foundation of good routing is a clean, current location master, exactly the kind of master-data discipline that makes or breaks every other WMS capability. For the wider role of the system of record here, see what is a WMS, and remember that a routing module is only ever as good as the WMS data underneath it.

The payoff: routing is the rare warehouse optimization that returns quickly and cheaply because it changes decisions, not concrete. Get the location master accurate, turn on a competent routing method inside the WMS, pair it with batching and periodic slotting, and coordinate across pickers to avoid self-inflicted congestion. Do that and you reclaim a meaningful slice of the shift that used to be spent walking the long way round. For where this sits in the full stack, keep the warehouse automation guide close.

8. References

The mathematical core of pick-path optimization is the travelling salesman problem, one of the most-studied problems in operations research: given a set of points that must all be visited and a cost of travel between each pair, find the visiting order that minimises total cost. Warehouse routing is a constrained, grid-geometry variant of it, and the practical solvers described above, nearest-neighbour construction, two-opt and or-opt local search, savings algorithms, simulated annealing and genetic metaheuristics, are the standard toolkit the operations research literature has developed for it over decades. The specific warehouse heuristics named in this article, the S-shape (serpentine), return and midpoint methods, are the classical routing policies documented in the warehousing and order-picking research literature. Readers who want to go deeper should search those named concepts, travelling salesman problem, order-picker routing policies, and warehouse pick-path optimization, in a reputable operations research or logistics reference rather than trusting a single vendor's description, since the field is mature, well documented and vendor-independent.

Final thoughts

Route optimization is the most under-appreciated win in warehouse operations precisely because it is not glamorous. There is no robot to photograph, no conveyor to cut a ribbon on, no dramatic before-and-after of a re-built floor. There is only software making a better decision about the order in which a person walks, and yet on a large-aisle, high-volume operation that decision reclaims a slice of labour that no amount of extra hiring would recover, because the extra hires would just walk the long way round too. It is the first optimization I would reach for, ahead of any capital project, on the simple grounds that it costs the least and returns the fastest.

The honest version of the pitch is the one worth remembering. Routing helps most when travel is a large share of the shift, when the location master is accurate, when it is paired with batching and slotting rather than run alone, and when it coordinates across pickers instead of optimising each in a vacuum. It hits a ceiling set by the physical layout, it is defeated by congestion if it ignores the other people on the floor, and it is only as good as the data in the WMS underneath it. Understand those limits and route optimization delivers exactly what it promises: less walking, more picking, and a warehouse that moves the same work with meaningfully less effort. Overstate it and you join the operations with a clever solver and a floor that still does not flow.

Related reading: Warehouse automation: the complete guide, AI for picking optimization, AI slotting optimization, AI workforce route & dispatch optimization, What is a WMS.