Recycling Sorting Robot Powered by Google Coral Edge TPU

Did you know that contamination rates in recycling facilities can reach up to 25%? That means one in four items never makes it into the recycling stream, largely due to human error during sorting. As climate change intensifies, reducing contamination is more critical than ever. Our solution—an autonomous sorting robot that uses Google Coral Edge TPU and machine‑learning object detection—cuts contamination rates dramatically while lowering operational costs.

Demo

Code Repository

Clone the full project on GitHub to follow along.

Step 1: Acquire the Dataset

Training an accurate object‑detection model requires a diverse set of labeled images. We used the TrashNet dataset, which contains 2,527 images across six classes:

• Glass – 501 images
• Paper – 594 images
• Cardboard – 403 images
• Plastic – 482 images
• Metal – 410 images
• Trash (non‑recyclable) – 137 images

The dataset is available for download on Google Drive. For quick training, download the pre‑resized archive (dataset-resized.zip). If you prefer custom resizing, use dataset-original.zip.

Step 2: Label the Images

We used labelImg to annotate bounding boxes around each object. Bounding boxes should snugly fit each item to maximize model accuracy. All annotation files (.xml) are stored in a dedicated folder. If labeling is time‑consuming, download our pre‑annotated dataset here.

Step 3: Train the Model

We employed TensorFlow’s transfer‑learning pipeline with the MobileNet‑SSD v2 (COCO) backbone, which is fully compatible with the Coral Edge TPU. Training can be performed locally or in the cloud. Local training is straightforward but slower; for accelerated results, use Google Cloud TPU or AWS GPU instances following Google’s tutorial or the EdjeElectronics guide.

Step 4: Compile for the Coral Edge TPU

After training, export the model as a frozen graph (.pb), then convert it to TensorFlow Lite with post‑training quantization. Finally, compile the TFLite model to an Edge TPU binary using edgetpu_compiler. For full instructions, see the Edge TPU compiler documentation.

Step 5: Deploy on Raspberry Pi

Set up a Raspberry Pi 4 with the Coral USB Accelerator and the Raspberry Pi Camera Module following the official setup guide. Once ready, run:

python test_detection.py \
  --model recycle_ssd_mobilenet_v2_quantized_300x300_coco_2019_01_03/detect_edgetpu.tflite \
  --labels recycle_ssd_mobilenet_v2_quantized_300x300_coco_2019_01_03/labels.txt

A live window will display detections; press q to exit.

Step 6: Build the Robotic Arm

The arm is a 3D‑printed, six‑degree‑of‑freedom manipulator sourced from Thingiverse. Follow the assembly instructions and connect the servos in the following order (bottom to top): 3, 11, 10, 9, 6, 5. Upload basicMovement.ino to the Arduino to test grabbing and placing.

Step 7: Integrate RPi and Arm

Mount the camera at the base of the arm’s claw to minimize perspective distortion. Upload roboticArm.ino to the Arduino, then connect the Raspberry Pi to the Arduino via USB to establish serial communication. Configure the serial port in the Arduino code to match the Pi’s /dev/ttyUSB0 interface.

Step 8: Final Touches

Encapsulate the electronics in a custom project enclosure for safety and aesthetics. A finished build looks like this:

Pre‑Trained Model

If you prefer to skip training, download our ready‑to‑deploy model here.

By leveraging machine learning on the Coral Edge TPU, this robot delivers precise, low‑contamination sorting at a fraction of the cost of human labor.