AI-Driven Adaptive Torque Allocation for Electric Vehicles on Rough Terrain 🚗⚡🌍

---

🔥 Project Overview

This project, selected among the finalists of the 8th edition of TECHgium (L&T Technology Services) from over 40,000+ students, proposes a novel AI-based torque allocation system for electric vehicles navigating rough terrains. The aim is to maximize vehicle stability, avoid skidding, and adaptively balance torque across all four wheels based on real-time sensor input.

We built a complete simulation using Gazebo, developed an optimization algorithm (SLSQP) to compute adaptive torque distribution, trained a deep learning model, and finally deployed it in a Raspberry Pi-powered prototype. The real-time control and monitoring were achieved via a custom-built Flask web application.

---

📁 Project Structure

Rubicon/
├── Evata/                          # EV Car Model
│   ├── Materials/
│   ├── Meshes/
│   ├── Thumbnails/
│   └── model.sdf
│
├── World_Assets/                  # Terrain Assets for Gazebo
│   ├── Materials/
│   ├── Meshes/
│   ├── Thumbnails/
│   └── model.sdf
│
├── try_world.sdf                  # Custom rough terrain Gazebo world

├── imu_reader5.py                 # IMU sensor reading and preprocessing script
├── training_data.csv              # Raw data from Gazebo simulation
├── ai_model.csv                   # Processed dataset after optimization and feature engineering
├── ai_model.ipynb                 # AI training notebook
├── ai_model.keras                 # Trained neural network model (exported)
├── index.html                     # Flask dashboard template
├── raspi3.py                      # Main hardware control script (Raspberry Pi + sensors + AI + Flask)

---

🚀 Key Highlights

🛠️ Phase 1: Simulation in Gazebo

• Rough Terrain World: Created a custom uneven terrain in Gazebo.
• EV Model: Integrated a 4-wheel-drive electric car model with sensor mounts.
• Data Logging: Captured ax, ay, az, wx, wy, wz from the IMU during runs.

📊 Phase 2: Data Processing & Optimization

• Velocity Calculation: Derived vx, vy using acceleration data.
• Wheel Speed Estimation: Applied rigid body kinematics.
• Slip Angle: Estimated slip angle for each wheel — a major factor in torque gaps.
• Optimization Algorithm: Used SLSQP to:
  • Ensure torque ≤ motor max torque.
  • Fulfill driver-demanded torque.
  • Balance left vs. right wheels.
• Torque to Duty Cycle: Converted optimized torque to RPM → Duty Cycle (for motor drivers).

🧠 Phase 3: AI Model Development

• Input Features: ax, ay, wx, wy, wz
• Target: fl_duty, fr_duty, rl_duty, rr_duty
• Model: Neural Network (4 hidden layers, ReLU + Dropout)
• Optimizer: Adam
• Output: .keras file + scaler.pkl

💻 Phase 4: Deployment on Raspberry Pi 4B

• Created a virtual environment for isolated dependency management.
• Installed packages like keras, scikit-learn, flask, RPi.GPIO, etc.
• Uploaded model and scaler to run real-time inference from IMU.

🔧 Phase 5: Hardware Prototype

• Chassis: 4-wheel drive (4 DC motors).
• Motor Control: Two L298N motor drivers for independent control.
• Sensors:
  • IMU: Real-time terrain sensing.
  • Ultrasonic Sensor: Obstacle avoidance (20cm threshold).
  • Buzzer: Alerts on obstacle detection.
• AI Inference: AI model predicts optimal duty cycles in real-time.

🌐 Phase 6: Flask-Based Web App (RC + Visualization)

• Control: Forward, Backward, Left, Right
• Live Graphs:
  • Roll vs. Time 📈
  • Pitch vs. Time 📉
  • Duty Cycle Distribution
• Toggle AI Mode ON/OFF to compare stability
• Emergency Stop Button
• Responsive and clean UI

---

🧠 Project Architecture

---

📽️ Project at a Glance

---

🌟 Results

Condition	Stability (Pitch/Roll)	Torque Distribution
Without AI	High Variance	Random/Fixed
With AI	Smoother Performance	Dynamic

• Significant improvement in vehicle stability
• Real-time torque balancing with AI
• Effective slip angle adaptation

---

🔮 Future Enhancements

• Integrate 360° LIDAR for advanced obstacle detection.
• Implement Reinforcement Learning for adaptive control.
• Expand model to handle multi-terrain transitions.
• Deploy on real EV platforms for field testing.
• Add cloud sync and mobile app control.

---

📦 Requirements

🧪 Software Setup

Install Python dependencies using:

pip install -r requirements.txt

This includes:

• tensorflow / keras for AI model inference
• scikit-learn for data preprocessing (StandardScaler)
• flask for the web dashboard
• matplotlib, numpy, etc.

---

🌍 Gazebo Harmonic (For Mac M1 Users)

If you are using Mac M1, follow these steps to run the simulation and collect IMU data:

1. Start the Gazebo Server:

   gz sim -v 4 try_world.sdf -s

2. Launch the Gazebo GUI (in a new terminal):

   gz sim -v 4 -g

3. To view IMU data from the running simulation, use:

   gz topic -e -t /imu

Ensure you have the EV model and terrain world files placed correctly in your project directory (try_world.sdf, Evata/, and World_Assets/).

---

🔩 Hardware Prototyping Requirements

To replicate the real-world testing on a prototype, you'll need:

Component	Quantity	Purpose
Raspberry Pi 4B	1	Main controller for AI inference + sensors
L298N Motor Driver	2	For independent control of 4 DC motors
DC Motors	4	To drive each wheel
MPU6050 Sensor	1	For real-time pitch, roll, and acceleration
Ultrasonic Sensor	1	For obstacle detection
Buzzer	1	Alerts when obstacle is near
Jumper Wires	Many	To connect components on breadboard
Li-Po Battery Pack	1	Power supply for motors + Raspberry Pi
Car Chassis	1	Mounting platform for motors and sensors
Breadboard	1	Prototyping connections
Power Bank	1 (optional)	Power Raspberry Pi independently

📦 Total POC cost ≈ ₹18,800 (as per Techgium presentation)

---

🤝 Connect With Me

Feel free to reach out if you're curious about:

• 🔍 AI/ML for real-world robotics
• 🤖 AI in embedded hardware systems
• 📈 Data science or computer vision applications
• 💡 Startup ideas around intelligent mobility

📧 ujjwalrajbgis@gmail.com
🔗 LinkedIn

🧑‍💻 Authors

• Ujjwal Raj – System Design, AI Modeling, Hardware & Integration

  Finalist at TECHgium 8th Edition | IIITDM Kancheepuram

---