I've opened some of my repos related to how I run YOLO object detectors on my OAK-D and Pi 5/AI Hat hardware

[Michael Wimble]

I've just opened a repo that might be useful if you're running computer-vision-based object detection on your robot: wimblerobotics/sigyn_ai

---

What it does

It's a complete, scripted pipeline that takes you from raw images all the way to a running detector on your robot:

Capture images → RoboFlow dataset → Train on GPU (or Colab) → Export → Compile → Deploy

No Jupyter notebooks required for the normal path — everything is shell scripts and YAML configs.

---

Who might find it useful

• Anyone running a Raspberry Pi 5 + Hailo-8 AI HAT or a Luxonis OAK-D / OAK-D Lite and wanting to train a custom detector against their own objects
• People who want a repeatable, version-controlled training workflow rather than ad-hoc scripts
• Beginners to YOLO/RoboFlow — the new ROBOFLOW_WORKFLOW.md guide walks through the whole RoboFlow process step by step, including click-by-click dataset settings for each device type

---

Tested hardware and models

Device	Model	Input	Performance
Pi 5 + Hailo-8 AI HAT	YOLOv8n INT8	512×512	~25–30 FPS
OAK-D / OAK-D Lite	YOLOv5n FP16	416×416	~18–25 FPS

Training a single-class detector (my Coke Zero cans, ~360 images) takes 5–10 minutes on a 3060. A Google Colab path is also documented for those without a local GPU.

---

One-command pipelines

# Pi 5 + Hailo-8: download dataset, train, export, compile .hef, deploy
./scripts/train_pi5_hailo.sh -v <roboflow_version> -n <run_name>

# OAK-D: download dataset, train, export, compile .blob, deploy
./scripts/train_oakd.sh -v <roboflow_version> -n <run_name>

• 
  
• 
  
• 
  
• 
  

---

Related repos (also being opened)

The full capture-to-train loop touches a few other repos I'm also making public:

• wimblerobotics/pi_can_detector — the Pi 5 + Hailo-8 ROS 2 detector node (subscribes to a joystick button topic to snapshot training images directly from the running robot)
• wimblerobotics/sigyn_oakd_detection — the OAK-D detector node
• wimblerobotics/sigyn_bluetooth_joystick — the joystick node that publishes the snapshot trigger

You don't need any of those to use sigyn_ai for training — they're only relevant if you want the on-robot image-capture workflow.

---

Apache 2.0. 

Let me know if things work for you, or could use improvements or clarification. Don't forget to start the repos if you like them. Don't forget to tip the wait staff.

These repos are intended to be starting points for your own work. I might make changes if they don't make major headaches in my usage on my robot.

— Michael