Video: https://drive.google.com/drive/folders/1q2jQ_tqfvA1bjxktAekHfnpSIeDnIRGD?usp=sharing

Inspiration

We've all done it. Something breaks at home, you drive to the hardware store, and then you stand in an aisle staring at forty kinds of screws with no idea which one you need. The staff are busy, so you guess, or you leave.

We looked it up and it's not just us. In Zebra's 2024 shopper study, 57% of shoppers left a store without something they came for, and 70% said it's hard to find an associate. Home Depot even ran a Kaggle competition in 2016, with more than 2,000 teams, just on matching what shoppers type to the right product. And a Hippo survey found 58% of DIY homeowners overspent because of mistakes.

Store robots have been tried before. Lowe's LoweBot could take you to a product if you knew its name. But our problem is that people don't know the name. So we started from the problem itself.

What it does

Aisle Guide is a kiosk, a page on your phone, and a guide robot.

You tap "I need help" on the kiosk and scan a QR code. No app to install. Then you just show it what's wrong: type it, take a photo, or record a voice note. It asks your budget, and gives you a short list of parts this store actually stocks, with a photo, price, aisle and bay for each.

Then the robot walks you there. When it gets within five metres, the shelf tag lights up and your phone turns green, so you know exactly where to look. Tap "I am done" and it takes you to the next item. At the end you get repair steps and how-to videos, and the robot heads back to its dock.

How we built it

n8n runs the whole thing. The kiosk, the phone and the robot each call a webhook, and every call is one short run: load the visit, decide, save, reply. Nothing sits waiting inside n8n.

It ended up bigger than we planned: 4 workflows, 132 nodes, 38 decision nodes and 7 webhooks. We named every decision node as a plain question, like "Relevant to home repair?" or "Budget known?", so we could read our own canvas at 2 a.m.

There are three AI calls. The first only checks that the request is about home repair. The second figures out what you need and picks items from the catalogue. The third writes the repair steps after the phone has already got its answer, so nobody waits for it. If a Gemini model is busy, the workflow tries the next one, with Claude as a backup.

We don't let the AI have the last word. It can only read the catalogue through a read-only database user, n8n re-checks every item it suggests, and n8n does the budget maths itself.

The rest: PostgreSQL for the catalogue and visits, a React phone app, and a robot simulated in ROS 2 with the real Nav2 navigation stack, shown in RViz. One command starts everything in Docker.

Challenges we ran into

  • Things happening at the same time. A double tap, a robot event and a timeout can all land in the same second. We had runs overwriting each other until we added a revision check, so only one save wins.
  • The AI making things up. It kept suggesting products the store doesn't sell. Checking every item against the catalogue fixed that.
  • Free-tier limits. Models went "busy" in the middle of testing, which is why the fallback chain exists.
  • Short answers. "About $30" looks off-topic on its own, so we had to show the relevancy AI the question it was answering.
  • Robot timing. Robot orders have to go out one at a time and in order, or the light and the phone get out of sync.

What we learned

The big one: keep the decisions in the workflow, not in the model. The AI gives advice, n8n decides. Once we did that, everything got easier to debug, because every run shows exactly which branch it took and why.

We also learned that short runs with the state kept in the database are far easier to live with than one long flow that waits.

Does it save anything?

Here's our rough estimate. From Home Depot's FY2023 10-K, a store sees about 11,380 in-store transactions a week. If 5% of those involve a "what do I need, where is it" question, each takes 2 minutes of staff time, and the phone answers half of them:

$$11{,}380 \times 0.05 \times 2\ \text{min} \times 0.5 \approx 9.5\ \text{hours per store per week}$$

At the US median retail wage of \$17.03 an hour (BLS), that's about \$8,400 per store per year of staff time freed up. The 5%, the 2 minutes and the half are our guesses, not measured numbers.

What's next

A real robot and real shelf lights, more than one robot per store, and a live catalogue with stock levels.

Built With

Share this project:

Updates

Submission history