How warehouse robots avoid getting in each other’s way

Imagine two warehouse robots approaching opposite ends of a narrow aisle. Each has a perfectly sensible route to its next job. Together, those routes may be impossible to follow. Making a robot navigate well is only part of the challenge: a working fleet must decide who uses a shared space, when, and where everyone else should wait.

This is the territory of multi-agent pathfinding. In its simplest research form, the floor becomes a network of positions and connections. A plan assigns each robot both a place and a time; waiting is an action too. That time dimension is what lets several machines use the same patch of floor without trying to occupy it together.[1]

An aisle needs a schedule as well as a map

A collision-free plan can still be inefficient. One robot might wait safely for a long time while several others take unnecessary detours. Pathfinding researchers therefore distinguish objectives such as finishing the last task sooner and reducing the sum of individual completion times. Those objectives can favour different plans for the same fleet.[1]

Now imagine a robot that delivers its load and parks at the entrance to the only usable aisle. Its own job is complete; everyone else’s route is blocked. This is why the destination matters beyond the moment of arrival. In research on continuing pickup-and-delivery tasks, Hang Ma and colleagues make available waiting places part of the conditions under which their method can keep tasks moving.[2]

Their guarantee depends on the layout: among other conditions, robots need enough non-task endpoints, with paths between endpoints that do not pass through other endpoints. The practical lesson is that suitable parking space belongs in the design of the system. A clever route planner cannot manufacture somewhere to wait when the floor offers nowhere suitable.[2]

Real robots cannot change direction like dots on a screen

A 2024 NAVER LABS research paper makes the physical complications explicit. Large warehouse robots can take substantial time and distance to accelerate or brake. Their speed can depend on the load. They occupy an area, rather than a mathematical point, and narrow aisles may prevent them from turning around. Some pickups also require approaching with a particular orientation.[3]

Consequently, a slot in the aisle must be long enough for the actual movement. Scheduling two robots a moment apart is no solution if the first is still braking, turning or collecting a load when the second arrives. The route, the timing and the physical job have to agree.[3]

The best journey may be one the system avoids

Coordination starts before a robot moves. In a 2022 account of its fulfilment operations, Amazon described choosing which storage pods to bring to workers and grouping work so that a visit supplies more required items. That can reduce the number of journeys competing for the same floor. Adding robots without addressing those journeys can increase congestion instead of removing the bottleneck.[4]

Consider two orders needing items from the same mobile shelf. If the work can be grouped, one visit may serve both. The benefit reaches beyond the robot assigned to that shelf: it also leaves room for other robots. This is an illustrative consequence of batching, rather than a measured result for a particular warehouse.[4]

AI can forecast the traffic before it forms

Planning also benefits from an estimate of future congestion. Amazon’s 2023 research combined recent robot movements and intended routes to predict extra travel time beyond unobstructed movement. Its model estimated where robots would probably be, then used those occupancy patterns to forecast delay. In simulations of two layouts using real trajectory data, the researchers reported a 4.4% throughput improvement over their comparison system.[5]

That result describes a simulation study. Its useful idea is broader: the fastest-looking route on an empty map can become a slow route once other planned journeys are included. A congestion forecast gives the planner a reason to choose differently before the queue has formed.[5]

Amazon’s DeepFleet work develops this approach using models trained on robot activity. In its August 2025 announcement, the company explained the attraction of a learned forecast: repeatedly simulating every interaction in a large fleet requires computation that the operating system also needs for its ongoing work. Predicting the next stretch of activity can provide a less expensive planning aid.[6]

The accompanying paper tests 60-second forecasts using seven days of data from seven floors held out of training. One successful design represents an individual robot alongside nearby robots and map features. It gives the model information about local interactions, rather than treating the whole warehouse only as a flat picture. The compared models differ in size and training data, so this does not establish that one representation always wins.[7]

Amazon also reported a 10% improvement in fleet efficiency in its announcement. That company claim is separate from the paper’s forecast evaluation: predicting congestion accurately and increasing completed warehouse work are related measurements, but they are not interchangeable. The demonstrated role described here is helping a planner anticipate congestion and coordinate the fleet’s work.[6]

The resulting picture is surprisingly practical. A fleet needs somewhere to wait, realistic travel times, sensible combinations of jobs and advance warning of crowded routes. The valuable intelligence lies in making those pieces cooperate. A robot taking a slightly longer path can be doing exactly the right thing if it helps the whole floor finish more work.

Report an error or suggest a correction