🧬 BIO-INSPIRED MICRO-MOBILITY: Autonomous Pod

High-Speed Adaptive Navigation & Affective HRI on Edge Platforms

Research Project: Development of a resource-constrained autonomous navigation system combining F1-style control, social robotics (WALL-E style), and generative design.

---

📖 Abstract

This project addresses the trade-off between speed and passenger comfort in modern service robots. We propose "Mini Cyber-Rickshaw", a micro-scale (1:28) autonomous architecture that leverages End-to-End Deep Reinforcement Learning to achieve high-speed navigation while maintaining social compliance. The system is designed to operate on low-cost edge devices (Raspberry Pi Zero 2 W) using a Sim-to-Real transfer pipeline.

Key Research Question: Can we deploy F1-grade racing algorithms on a $15 computer to serve smart city tourism?

🌟 Key Features

1. 🏎️ F1-Inspired Navigation (The Brain)

• Algorithm: Proximal Policy Optimization (PPO) with MobileNetV3 backbone.
• Behavior: Optimized for Apex Clipping (cutting corners) and dynamic velocity control based on track curvature.
• Reward Function: Custom multi-objective function balancing Velocity, Cross-Track Error, and Jerk (smoothness).

2. 🤖 Active Perception & HRI (The Soul)

• Active Vision: Pan-Tilt camera mechanism (2-DOF) simulating saccadic eye movements to expand Field of View (FoV) at intersections.
• Affective Locomotion: State-machine based behaviors expressing internal states (e.g., Curiosity at landmarks, Fear at obstacles).

3. 🦴 Generative Bio-Chassis (The Body)

• Design: Voronoi lattice structure generated via Topology Optimization (Fusion 360).
• Performance: 40% weight reduction compared to stock chassis, maintaining structural integrity for high-speed impacts.

---

🛠️ System Architecture

Component	Specification
Platform	WLToys K989 (1:28 Scale RC Car)
Compute	Raspberry Pi Zero 2 W (Quad-core ARM Cortex-A53)
Sensors	Pi Camera V2 (160° FoV)
Actuation	PCA9685 PWM Driver + SG90 Servos (Head)
Simulation	Unity 3D / Donkey Car Simulator

---

🚀 Installation & Setup

1. Prerequisites

• Anaconda or Miniconda
• Python 3.9+
• NVIDIA GPU (Recommended for training)

2. Environment Setup

# Clone the repository
git clone [https://github.com/minhquang0407/mini-cyber-rickshaw.git](https://github.com/minhquang0407/mini-cyber-rickshaw.git)
cd mini-cyber-rickshaw

# Create Conda environment
conda create -n donkey python=3.9 -y
conda activate donkey

# Install dependencies
pip install -e .[pc]
pip install -r requirements.txt

3. Launch Simulator (Sim Mode)

Ensure myconfig.py is configured with DONKEY_GYM = True.

python manage.py drive

Access the web controller at: http://localhost:8887

---

🧠 Training Pipeline (Sim-to-Real)

We utilize a Behavioral Cloning approach augmented with DRL:

1. Data Collection: Drive manually in the Simulator to generate ~10,000 samples.
2. Training:

   donkey train --tub ./data --model ./models/mypilot.h5

3. Deploy: Transfer the .h5 model to Raspberry Pi Zero.
4. Inference:

   python manage.py drive --model ./models/mypilot.h5

---

📂 Project Structure

mini-cyber-rickshaw/
├── data/               # Raw training data (Git ignored)
├── docs/               # Research papers and diagrams
├── models/             # Trained Neural Networks
├── src/                # Custom source code (RL agents, Vision)
├── manage.py           # Main entry point
└── myconfig.py         # Configuration file

---

📊 Preliminary Results

(Placeholders - Update with your real charts)

• Lap Time: Reduced by 15% compared to PID controller.
• Inference Speed: 22 FPS on Raspberry Pi Zero 2 W.
• Smoothness: Jerk metric reduced by 30% using the proposed Reward Function.

---

👨‍💻 Author

• NGUYEN MINH QUANG
• Researcher
• University of Science - VNU
• Email: minhquang04072005@gmail.com

🤝 Acknowledgments

Special thanks to the Donkey Car Community and Stable Baselines3 team for the open-source tools.