GridNav-AI 🤖🗺️

A robot navigation project exploring two fundamentally different approaches to the same problem — Supervised Learning (ResNet CNN imitating BFS) and Reinforcement Learning (DQN-LSTM learning through trial and error).

---

Demo

Supervised Learning (ResNet)

Reinforcement Learning (DQN-LSTM)

---

Two Approaches, One Problem

	Supervised (ResNet)	Reinforcement (DQN-LSTM)
How it learns	Imitates BFS optimal paths	Trial and error with rewards
Needs labels?	✅ Yes — BFS solutions	❌ No
State input	Full grid (3-channel tensor)	5×5 vision window
Actions	8 directions	4 directions
Architecture	ResNet CNN	DQN-LSTM
Generalization	Any grid same size	✅ Random grids, variable density
Training time	~10-15 min	~30-60 min

---

How Each Approach Works

Supervised Learning Pipeline

1. Random Grid Generation — generates grids with obstacles, robot R, and target T
2. BFS Optimal Pathfinding — finds shortest path (used as training labels)
3. Dataset Creation — each BFS step becomes a (state, action) training pair
4. ResNet Training — CNN learns to predict next optimal move from grid state
5. Simulation — trained model navigates new unseen grids

Reinforcement Learning Pipeline

1. Random Grid Generation — new random grid every episode, variable obstacle density
2. DQN-LSTM Agent — 5×5 vision window, hidden state carries memory between steps
3. Reward Shaping — warm/cold signal toward target + revisit penalty
4. Fixed Eval Set — 20 fixed grids measure true generalization (like validation set)
5. Early Stopping — stops when 90% success rate achieved on eval set

---

Model Architectures

ResNet CNN (Supervised)

Input (3 channels):
├── Channel 0: Obstacle map
├── Channel 1: Robot position
└── Channel 2: Target position
        ↓
Initial Conv2D (32 filters) + BN + ReLU
        ↓
Residual Block 1 (32 → 32)
        ↓
Residual Block 2 (32 → 128, with 1×1 downsample)
        ↓
Flatten → Dropout(0.6) → FC(128) → Dropout(0.6) → FC(8)
        ↓
Output: 8 actions (UP, DOWN, LEFT, RIGHT, diagonals)

DQN-LSTM (Reinforcement)

Input (30 values):
├── 5×5 vision window (25 values)
│     0.0=free, 0.5=OOB, 1.0=obstacle, 3.0=target
├── Normalized robot position (2 values)
├── Normalized target position (2 values)
└── Exploration progress (1 value)
        ↓
Encoder: Linear(30→128) → LayerNorm → ReLU → Linear(128→128) → LayerNorm → ReLU
        ↓
LSTM(128→128) — carries memory (h,c) between steps
        ↓
Decoder: Linear(128→64) → ReLU → Linear(64→4)
        ↓
Output: 4 Q-values (UP, DOWN, LEFT, RIGHT)

---

Training Results

Supervised Learning

Metric	Value
Training Samples	5,000
Grid Size	15×15
Obstacle Density	25%
Epochs	50 (early stopping)
Optimizer	Adam (lr=5e-4, wd=5e-3)
Loss	CrossEntropyLoss

Reinforcement Learning

Metric	Value
Grid Size	15×15
Obstacle Density	10%–35% (random each episode)
Episodes	5000–8000
Optimizer	Adam (lr=1e-3, wd=1e-4)
Eval Set	20 fixed grids
Best Success Rate	~70% on unseen random grids

---

Project Structure

GridNav-AI/
├── README.md
├── requirements.txt
│
├── src/
│   ├── path_finder.py              # Supervised ResNet training + simulation
│   ├── reinforcement_lesson_2.py   # Q-Table (fixed grid)
│   ├── reinforcement_lesson_3.py   # DQN (position + target)
│   ├── reinforcement_lesson_4.py   # DQN blind robot (reward shaping)
│   ├── reinforcement_lesson_5.py   # DQN-LSTM (5×5 vision, random grids) ← main RL
│   └── reinforcement_lesson_6.py   # BPTT (sequence training, experimental)
│
├── demo/
│   ├── app.py                      # Streamlit home page
│   ├── core/
│   │   ├── grid_utils.py           # Shared grid generation + rendering
│   │   ├── rl_model.py             # DQN-LSTM inference + training
│   │   └── supervised_model.py     # ResNet inference + training
│   └── pages/
│       ├── 1_Training.py           # Live training (RL + Supervised)
│       ├── 2_Inference.py          # Side-by-side model comparison
│       └── 3_Grid_Builder.py       # Draw custom grids, benchmark models
│
├── models/
│   ├── stage3_best.pth             # Best RL model
│   └── supervised_best.pth         # Best supervised model
│
├── examples/
│   ├── robot_animation.gif
│   ├── stage3_animation.gif
│   ├── training_history.png
│   └── stage3_rewards.png
│
└── .gitignore

---

Requirements

pip install -r requirements.txt

torch>=2.0.0
numpy>=1.24.0
matplotlib>=3.7.0
tqdm>=4.65.0
pillow>=9.0.0
streamlit>=1.28.0
plotly>=5.17.0

---

Setup & Run

Clone

git clone https://github.com/WeskerPRO/GridNav-AI.git
cd GridNav-AI

Train Supervised (ResNet)

cd src
python path_finder.py

Train Reinforcement (DQN-LSTM)

cd src
python reinforcement_lesson_5.py

Run Streamlit Demo

cd demo
streamlit run app.py

The demo has three pages:

• Training — watch RL agent or ResNet train live with real-time curves
• Inference — load trained models, compare RL vs Supervised side by side
• Grid Builder — draw your own maze, benchmark all models + BFS

---

RL Training Notes

💡 Per-episode reward oscillates on random grids — this is expected. Grid difficulty varies each episode. Use success rate on the fixed eval set as the true learning metric (equivalent to validation accuracy).

⚠️ Train on the same grid size you test on. RL model normalizes positions by grid dimensions — a model trained on 15×15 expects 15×15 inputs.

🧠 The model needs the target coordinates in its state to generalize across random grids. Without them, reward shaping sends contradictory signals on different grid layouts.

---

Roadmap

✅ Completed

• ✅ Random grid generator with guaranteed solvable paths
• ✅ BFS optimal pathfinding for label generation
• ✅ ResNet CNN with residual blocks + Dropout regularization
• ✅ Training with early stopping + LR scheduler
• ✅ Q-Table navigation (fixed grid)
• ✅ DQN with position + target coordinates
• ✅ DQN blind robot with reward shaping
• ✅ DQN-LSTM with 5×5 vision window
• ✅ Random grid training for generalization
• ✅ Fixed eval set (validation equivalent for RL)
• ✅ Streamlit demo with live training + inference + grid builder
• ✅ BPTT sequence training (experimental)
• ✅ 3D grid pathfinding (experimental)

🚧 In Progress

• 🔄 Model generalization improvement (target: 80%+ success rate)

📋 Upcoming

• Fog of war exploration (partial map reveal)
• Larger grid support (35×35+)

---

License

This project is licensed under the MIT License — free to use, modify, and distribute.

---

Made with ❤️ by WeskerPRO

Supervised Learning meets Reinforcement Learning — same robot, two minds.