FitCoach-AI – Bicep Curl Form Coach

FitCoach-AI is a small end-to-end system that uses phone motion sensors (accelerometer + gyroscope) and a machine learning model to give real-time feedback on bicep curl form. It tells you whether a rep looks CORRECT or INCORRECT and keeps basic rep statistics.

⚠️ Scope: Right now this project only supports bicep curls. The architecture is designed so more exercises (squats, push-ups, etc.) can be added later with new models and UI updates.

---

Project structure

• train_bicep_model.py, classify_new_session.py
  • Python scripts to train a RandomForest classifier from IMU CSVs and classify new sessions into CORRECT vs INCORRECT windows.
• models/bicep_model.joblib
  • Trained model used by the backend API.
• server.py
  • FastAPI backend that exposes POST /classify_window and returns { "label": "CORRECT" | "INCORRECT", "confidence": float } for a window of IMU samples.
• android_app/
  • Android client (Kotlin) that reads accelerometer/gyroscope data, buffers it into sliding windows, sends them to the backend, and shows:
    • Live status (CORRECT / INCORRECT).
    • Total reps, correct reps, incorrect reps in a simple fitness-style UI.

---

How to run (local dev)

1. Backend (FastAPI + model)

   • Install Python dependencies.

   • Make sure models/bicep_model.joblib exists (run python train_bicep_model.py if needed).

   • Start the API:

     uvicorn server:app --host 0.0.0.0 --port 8000 --reload

   • Open http://127.0.0.1:8000/docs to see the API docs.

2. Android app (phone + sensors)

   • Open android_app in Android Studio.

   • Update the base URL in MainActivity.kt:

     private const val BASE_URL = "http://YOUR_LAPTOP_IP:8000"

     (Laptop and phone must be on the same Wi‑Fi network.)

   • Ensure network_security_config.xml allows HTTP to this IP for local development.

   • Run the app on a real Android device, tap Start, and perform bicep curls with the phone/watch on your training arm.

---

How it works

• The Android app collects IMU samples, keeps a sliding 3‑second window, and periodically sends a window to the backend when:
  • it has enough samples,
  • there is enough change in acceleration to look like a rep, and
  • at least ~1–1.5 seconds have passed since the last counted rep.
• The FastAPI service converts the window into features and calls the trained RandomForest model, which outputs CORRECT vs INCORRECT and a confidence score.
• The app updates:
  • Status text (CORRECT / INCORRECT + confidence).
  • Rep counters (total, correct, incorrect).

---

Future work

Planned extensions:

• Add more exercises (e.g., squats, push-ups) by training additional models or extending labels (BICEP_CORRECT, SQUAT_INCORRECT, etc.).
• Per-user calibration and better rep segmentation (per‑rep peak detection instead of simple sliding windows).
• More advanced UI with history, sets, and progress tracking similar to commercial fitness apps.

---