Object Detector Model

Context: During the URC competition we must be able to detect two types of objects: mallet and water bottle. These objects will exist at two distinct waypoints during the autonomous mission. For a full context on the auton mission please read the URC Rules, specifically 1.f Autonomous Navigation Mission.

Problem: The majority of the R&D went into the object detector's execution/inference environment, leaving the model left somewhat rudimentary. The current model is a YOLOv8 small with fine tuning using our own mallet and water bottle dataset and RoboFlow.

Solution: Use a smaller, lighter 🪶 model using one of:

• Tensor Flow:
  • Instructions to do
  • Model Bank
  • The smallest one I found was 9.4 MB
  • Convert .tflite to onnx
• YoloV8:
  • YoloV8n has a size of 6.4 MB
• Training Custom Model (Using YOLO)
  • Process

---

Interface (subject to change): Use the existing TensorRT framework to execute the model and then publish the objects into the tf tree. Here are some of the tf tags:

bottle
mallet
immediateMallet
immediateBottle