Raspberry Pi AI Kit with Hailo-8L: Edge Object Detection on a Budget 2026

TL;DR:

• The Raspberry Pi AI Kit adds a Hailo-8L M.2 NPU to the Pi 5, delivering 13 TOPS of neural network inference for around £65.
• It handles real-time object detection (COCO classes at 30fps), pose estimation, and instance segmentation without using the Pi’s CPU.
• The RPi software stack (rpicam-apps with Hailo post-processing) makes getting started straightforward; custom models require ONNX conversion.
• It’s not a replacement for Jetson Orin or Coral TPU for demanding production deployments, but it’s the most accessible entry point to dedicated edge AI hardware.

Most edge AI discussions focus on industrial-grade hardware — Jetson Orin modules, Coral TPU accelerators, or Qualcomm’s RB-series platforms. These are excellent for production deployments, but they cost £150–£500 and require more setup time than many prototyping or hobbyist use cases justify.

The Raspberry Pi AI Kit occupies a different niche: a £65 M.2 add-on that transforms a Pi 5 into a capable edge AI platform. It launched in mid-2024 and has matured considerably since. In 2026, with updated software support and a growing library of compatible models, it’s the easiest way to add real-time neural network inference to an edge device.

What’s in the Kit

The AI Kit consists of:

• Hailo-8L M.2 accelerator — an NPU (neural processing unit) from Hailo, designed specifically for neural network inference. The -8L variant delivers 13 TOPS (tera operations per second).
• M.2 HAT+ — the PCIe adapter board that connects the M.2 card to the Raspberry Pi 5’s PCIe FFC connector.

The Raspberry Pi 5 is not included. You need a Pi 5 (any RAM variant), the AI Kit, and a power supply that can sustain the combined draw. The official 27W USB-C power supply is recommended.

What It Actually Does

The Hailo-8L accelerates inference on neural networks that have been compiled for Hailo’s runtime. The official software stack focuses on computer vision:

Object detection: YOLOv5, YOLOv8, and YOLOX models run at 30fps on 1080p input, detecting people, vehicles, animals, and objects across the 80 COCO classes. This runs entirely on the Hailo chip, leaving the Pi 5’s CPU free for application logic.

Pose estimation: Identifies body keypoints (shoulders, elbows, knees, etc.) in real time, which powers gesture-based interfaces, exercise analysis, and safety monitoring.

Instance segmentation: Goes beyond bounding boxes to produce per-pixel masks for each detected object. Useful for robotic picking applications or precise spatial tracking.

Face detection and recognition: A face detection pipeline is included in the official examples. Note that face recognition for identification requires careful consideration of local privacy law.

Getting Started

Raspberry Pi OS Bookworm is required. The Hailo setup is handled through APT packages:

sudo apt update && sudo apt upgrade -y
sudo apt install hailo-all

This installs the Hailo PCIe driver, the HailoRT runtime, the TAPPAS framework, and the rpicam-apps integration. Reboot after installation.

Verify the Hailo device is recognised:

hailortcli fw-control identify

If this returns device information (firmware version, device ID), the hardware is working correctly.

Run the bundled object detection demo with a connected camera:

rpicam-hello -t 0 --post-process-file \
  /usr/share/rpi-camera-assets/hailo_yolov5_personvehicle.json \
  --lores-width 640 --lores-height 640

This opens a live preview window with bounding boxes drawn around detected persons and vehicles.

Working With Custom Models

The bundled models cover common use cases. For custom applications, you need to:

1. Train or obtain a model in ONNX format
2. Compile it for Hailo’s runtime using the Hailo Dataflow Compiler (available free for Hailo-8 family devices)
3. Deploy the compiled .hef file

The compilation step requires a Linux machine with the Dataflow Compiler installed (not the Pi itself). It’s more involved than deploying to a Coral TPU, but the Hailo developer documentation is thorough and their community Discord is active.

A practical shortcut: the Hailo Model Zoo on GitHub includes pre-compiled .hef files for dozens of architectures. If your use case maps to a model already in the zoo, you can skip compilation entirely.

Real-World Performance Numbers

From community benchmarks and our own testing:

These run without CPU involvement. The Pi 5’s CPU is simultaneously available for streaming results, triggering actions, running a web server, or doing other application work.

Power draw for the combined Pi 5 + AI Kit under inference load is approximately 8–10W — manageable for battery-powered field deployments with a decent USB power bank.

When It’s Not Enough

The 13 TOPS of the Hailo-8L has limits. For applications requiring multiple simultaneous detection pipelines, higher-resolution inputs, or model complexity beyond standard vision architectures, you’ll hit the ceiling quickly.

The Hailo-8 (the full version, 26 TOPS) addresses some of this, but requires a different form factor. For serious production deployments with reliability and thermal management requirements, Jetson Orin Nano (40 TOPS, active cooling, full CUDA stack) remains the professional choice.

The Pi AI Kit’s value is primarily in prototyping, education, and low-volume deployments where the £65 entry price and simplified setup justify the capability tradeoff.

Practical Use Cases

Where the Pi AI Kit makes sense in 2026:

• Retail analytics: people counting and dwell time analysis in small shops or pop-up locations, without sending video to the cloud
• Workshop safety: detecting whether PPE is being worn in small-scale manufacturing environments
• Garden monitoring: wildlife detection and classification without a cloud subscription
• Industrial inspection prototypes: testing detection models locally before moving to production hardware
• Education: the most accessible way to teach neural network inference to students and hobbyists

The combination of low cost, familiar Pi ecosystem, and genuine neural accelerator performance makes it a compelling starting point for anyone building towards edge AI deployment.