Warehouse Picking Route Optimization for Science Fair

ISEF Category: Engineering Technology: Statics and Dynamics

Subcategory: Industrial Engineering-Processing  ·  Difficulty: Advanced  ·  Setup: Home Setup  ·  Time: 1 to 2 Months

The Hook

Every extra step in a warehouse costs time and money. That means a tiny routing improvement can scale up fast. You can test whether a learned policy beats classic warehouse heuristics on real pick-lists. That gives you a clean mix of AI, optimization, and operations research.

What Is It?

This project asks a simple question with real business value, how should a picker move through a warehouse to collect items with the fewest wasted steps? You can think of it like planning the best grocery run, except the shelves are numbered, the list is longer, and every detour adds cost.

A reinforcement-learning agent learns by trial and error inside a simulated warehouse. A heuristic, like the S-shape rule, follows a fixed strategy that works well in many layouts. OR-Tools gives you a stronger optimization baseline, so you can compare learning against a classic algorithmic approach. In this project, you are not just building a model. You are testing whether the model makes better route decisions than established methods.

SKU means stock keeping unit. That is the unique code a store uses for each product. You can generate pick-lists from a public Amazon product dataset, then map those items into warehouse locations and measure how much travel each method requires.

Why This Is a Good Topic

This is a strong science fair topic because you can turn a real logistics problem into measurable numbers. You can compare route length, turn count, picking time proxy, and consistency across many random orders. That gives you clear evidence, not just a pretty demo. You also learn how simulation, optimization, and machine learning fit together, which is a useful skill set for engineering, data science, and supply chain work.

Research Questions

• How does a reinforcement-learning policy compare with the S-shape heuristic in total travel distance for random pick-lists?
• What is the effect of warehouse layout density on the gap between OR-Tools VRP and the heuristic baseline?
• Does training on one SKU distribution improve performance on a different SKU distribution?
• To what extent does order size change the advantage of reinforcement learning over fixed routing rules?
• Which reward design leads to the lowest average pick-route length across unseen orders?
• How does adding turn penalties affect route quality in the custom Gymnasium environment?

Basic Materials

• Laptop or desktop computer with at least 8 GB RAM.
• Python installed with Jupyter Notebook.
• Gymnasium for building the custom warehouse environment.
• OR-Tools for route optimization baseline tests.
• Pandas for cleaning SKU and order data.
• NumPy for array handling and simulation.
• Matplotlib or Seaborn for plotting results.
• Free Amazon product dataset or other open product catalog.
• Spreadsheet software for tracking trials and summaries.

Advanced Materials

• Laptop or desktop computer with a dedicated GPU, if available.
• Python with PyTorch or TensorFlow for reinforcement learning.
• Gymnasium for the custom environment.
• OR-Tools for baseline vehicle routing or picking path optimization.
• NetworkX for graph-based warehouse layout modeling.
• SciPy for statistical tests and confidence intervals.
• ImageJ or a plotting tool for layout visualization, if you create map-style figures.
• Large open product dataset for generating many SKU pick-lists.
• Version control software such as Git for tracking code changes.

Software & Tools

• Python: Runs the simulation, data processing, and optimization code.
• Jupyter Notebook: Lets you test ideas, inspect outputs, and document results in one place.
• Gymnasium: Provides the environment framework for training a reinforcement-learning agent.
• OR-Tools: Solves routing baselines that you can compare against your learned policy.
• Pandas: Cleans product data and turns orders into structured SKU lists.

Experiment Steps

1. Define the warehouse layout, the picker start point, and the exact path metric you will optimize.
2. Build a simple simulator that turns SKU lists into locations and calculates route cost.
3. Choose one baseline heuristic, one optimization method, and one reinforcement-learning agent to compare.
4. Decide how you will generate repeatable pick-lists from the public product dataset.
5. Design reward signals, state inputs, and action choices for the Gymnasium environment.
6. Plan your evaluation table so every method gets tested on the same unseen orders with the same metrics.

Common Pitfalls

• Mapping products to warehouse slots in a way that changes between runs, which makes your comparisons unfair.
• Comparing methods on different pick-list sets, which hides the real performance gap.
• Training the agent on the same orders used for testing, which inflates the results.
• Using route length alone and ignoring turns or backtracking, which can make a weak policy look better than it is.
• Building a reward that is too vague, which causes the reinforcement-learning agent to learn a shortcut that does not match warehouse efficiency.

What Makes This Competitive

A stronger version of this project does more than report one average distance score. It tests several warehouse layouts, several order sizes, and several reward designs, then checks whether the results hold on unseen data. You can make it stronger by using a clean train-test split, confidence intervals, and a fair baseline setup. A project like that shows real thinking about algorithm design, not just code that runs.

Project Variations

• Test the same routing methods on cold-storage warehouse layouts, where aisle access may change the best path.
• Compare total distance with turn count, to see whether the fastest route also makes the fewest direction changes.
• Swap the Amazon product catalog for a grocery order dataset, then measure whether item mix changes the best algorithm.

Learn More

• Google OR-Tools documentation: Find the vehicle routing section and the examples for path optimization and constraint solving.
• Gymnasium documentation: Read the custom environment guides to learn how reinforcement-learning simulations are structured.
• MIT OpenCourseWare, Introduction to Algorithms: Search for routing, graph, and optimization lectures that explain the logic behind baseline methods.
• PubMed: Search for review articles on warehouse order picking, route optimization, and supply chain efficiency.
• NOAA Data Discovery and Access or NASA Earthdata: Use these government data portals as examples of how to handle large public datasets cleanly, even though your topic is different.