Logistics ops · Production

Zalando trains a convolutional neural network to accelerate warehouse order batching

The problem

Zalando's OCaPi algorithm produced accurate optimal picker routes but took several seconds per pick list, making it too slow to use in real-time for batching the thousands of orders placed every hour.

First attempt

Directly combining OCaPi with simulated annealing to find near-optimal pick list splits was computationally infeasible because OCaPi required several seconds per pick list evaluation, which is too slow across thousands of pick lists.

Workflow diagram · grounded in source
1
Order intake trigger
trigger
“thousands of new orders are placed every hour, and each order must be assigned to a pick list”
2
Neural network travel time estimation
ai_action
“we generated 1 million random pick lists, and used OCaPi to give each list a "label": the calculated travel time. Then we fed the coordinates of the pick lists along with the travel times into a convolutional neural network”
3
Simulated annealing batch optimization
routing
“a very simple optimization algorithm based on simulated annealing that starts with 40 orders of 2 items, each split randomly between two pick lists”
4
Optimized pick list output
output
“optimized batches allowed the workers to decrease their travel time per item picked by an average of 11%, compared with a random batch”
Reported outcome

A convolutional neural network trained as a fast OCaPi surrogate achieved 0.895% average estimation error, enabling simulated annealing to optimize order batching and reduce worker travel time by 11% versus random batching in an experimental setting.

Reported metrics
Neural network average relative error0.895%
Neural network absolute travel time error per hour32.25 seconds per hour
GPU training speedup over CPUfactor of 20
GPU training time for 1 million examples52.6 seconds
Show all 5 reported metrics
neural network average relative error0.895%
neural network absolute travel time error per hour32.25 seconds per hour
GPU training speedup over CPUfactor of 20
GPU training time for 1 million examples52.6 seconds
worker travel time reduction vs random batching (experimental)11%
Reported stack
CaffecuDNN_v2OpenBLASNVIDIA Tesla K80
Source
https://engineering.zalando.com/posts/2015/12/accelerating-warehouse-operations-with-neural-networks.html
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

A convolutional neural network trained as a fast OCaPi surrogate achieved 0.895% average estimation error, enabling simulated annealing to optimize order batching and reduce worker travel time by 11% versus random bat…

What tools did this team use?

Caffe, cuDNN_v2, OpenBLAS, NVIDIA Tesla K80.

What results were reported?

Neural network average relative error: 0.895%; Neural network absolute travel time error per hour: 32.25 seconds per hour; GPU training speedup over CPU: factor of 20; GPU training time for 1 million examples: 52.6 seconds (source-reported, not independently verified).

What failed first in this deployment?

Directly combining OCaPi with simulated annealing to find near-optimal pick list splits was computationally infeasible because OCaPi required several seconds per pick list evaluation, which is too slow across thousand…

How is this logistics ops AI workflow structured?

Order intake trigger → Neural network travel time estimation → Simulated annealing batch optimization → Optimized pick list output.

Related logistics ops cases