🧠 Brawl Stars Gadget Detection Bot

This project is a real-time object detection bot for Brawl Stars, powered by Ultralytics YOLOv8.
It detects UI elements such as when a gadget is charged and automatically simulates a tap using ADB on your Android device.

🎮 What It Does

• Mirrors your game screen using scrcpy
• Captures the screen using mss in real-time
• Runs YOLOv8 on each frame to detect objects like:
  • gadget_charged
  • gadget_not_charged
• Sends tap commands using pure Python ADB integration when actions are triggered

---

🧩 Components

project/
├── gadget.py         # Tap coordinates for gadgets
├── test.py           # Main detector and action loop
├── best.pt           # Trained YOLOv8 model
├── data.yaml         # Dataset configuration file
├── dataset/
│   ├── train/
│   │   ├── images/
│   │   └── labels/
│   └── val/
│       ├── images/
│       └── labels/

---

⚙️ How to Train a Model

1. Prepare Your Data

Use a tool like makesense.ai or CVAT to annotate screenshots from Brawl Stars.

You must export using:

✅ YOLO with Images

Structure:

dataset/
├── train/
│   ├── images/
│   └── labels/
├── val/
│   ├── images/
│   └── labels/
data.yaml

2. Create data.yaml

train: dataset/train/images
val: dataset/val/images
nc: 2
names: ['gadget_charged', 'gadget_not_charged']

3. Train the Model

yolo detect train data=data.yaml model=yolov8n.pt epochs=30 imgsz=640 device=0

---

❌ What Went Wrong

❗ Labeling Took Too Long

The project didn’t progress as planned because the manual labeling process was too slow and time-consuming. While the tools worked correctly (like makesense.ai), the effort required to label enough examples for training made it difficult to build a large, high-quality dataset.

Why It Failed:

• Labeling each image manually became tedious and repetitive
• Not enough labeled data to train a reliable YOLOv8 model
• Dataset remained too small for good detection performance

🕐 Why It Didn’t Work Out

Even after organizing the dataset and setting up training, the combination of:

• Manual labeling overhead
• Limited compute power (GTX 1650 GPU)
• Small dataset

...led to slow progress. This blocked testing and real-time integration into the bot.

Possible solutions:

• Use Google Colab for training
• Use fewer epochs or a smaller model (yolov8n)
• Collaborate to speed up labeling

---

🚀 Future Ideas

• Add detection for enemies, power cubes, and supers
• Implement smarter action logic (e.g., auto-aim)
• Use segmentation or tracking models for more precise control

---

🙌 Credits

• Ultralytics YOLOv8
• scrcpy for screen mirroring
• mss for screen capture
• ADB for Android input simulation
• Chatgpt for making this :)