Custom Petoi vision model

Use this guide to build, train, and deploy a custom Petoi vision model.

Getting started

• Object detection - Explain the basics of object detection.
• How We Teach a Robot to "See" — Explains datasets, models, SenseCraft, and the end-to-end vision workflow.

Steps on how to get started

• Step 1: Choose or make your own dataset

  • Choose a pre-labeled dataset - Finds, verifies, and downloads pre-labeled object-detection datasets using Roboflow or Kaggle.
  • Creating custom training datasets — Creates, labels, and exports object-detection datasets with Label Studio.

    • Mac — Installs and starts Label Studio on macOS with Homebrew.
    • Windows — Installs and starts Label Studio on Windows.
• Step 2: Train Model

  • Train model on Google Colab - Loads datasets, trains models, and exports quantized files using Google Colab.
  • Training models on the cloud — Train YOLOv8 in Google Colab. Exports and Vela optimize a TFLite model.
  • Train models locally on Windows — Connects Colab to a local Jupyter runtime and GPU. Keeps training files on Windows.
• Step 3: Deploy the model

  • Deploy model to SenseCraft — Uploads a trained model to SenseCraft and tests it with the Petoi camera.

Workflow

This guide covers the complete workflow:

1

Create a dataset

Collect images and draw bounding boxes around each target object.

2

Train a model

Train a YOLOv8 object-detection model locally or in Google Colab.

3

Optimize the model

Export the model to quantized TFLite and optimize it with Vela.

4

Deploy the model

Upload the model to Grove Vision AI V2 through SenseCraft.

5

Connect the robot

Use detection results to trigger Petoi robot actions