Amazon Robotics Route Optimization with Machine Learning

A hybrid optimization and machine learning project that simulates warehouse route optimization for Amazon Robotics, combining traditional operations research (OR-Tools) with ML-based travel time prediction to improve robot routing efficiency.

Project Overview

This project demonstrates a complete pipeline for warehouse route optimization, simulating a mini Amazon warehouse where robots must:

• Collect packages from pick-up points
• Deliver them to specific delivery stations
• Choose optimal routes to minimize travel time
• Adapt to congestion and dynamic warehouse conditions

The solution integrates:

1. Warehouse Simulation: 2D grid-based warehouse with robots, pick-up points, and delivery stations
2. Route Optimization: Google OR-Tools for solving TSP (Traveling Salesman Problem) and VRP (Vehicle Routing Problem)
3. Machine Learning: scikit-learn models to predict travel times based on congestion, obstacles, and warehouse conditions
4. Hybrid Approach: ML-enhanced optimization that uses predicted travel times instead of simple distances

Key Features

• 2D Warehouse Simulation: Configurable grid-based warehouse with realistic constraints
• OR-Tools Integration: Efficient TSP/VRP solving with multiple optimization strategies
• ML-Based Prediction: Random Forest and MLP models to predict travel times under various conditions
• Comprehensive Visualizations: Route comparisons, error analysis, and congestion heatmaps
• Reproducible Pipeline: Complete workflow from data generation to final results

Repository Structure

amazon-robotics-route-optimization-ml/
│
├── src/
│   ├── simulate_warehouse.py       # Warehouse simulation and distance matrix generation
│   ├── optimize_routes.py          # OR-Tools route optimization (TSP/VRP)
│   ├── generate_ml_data.py         # ML dataset generation
│   └── utils.py                    # Helper functions (distances, plotting, etc.)
│
├── notebooks/
│   ├── 03_ml_data_generation.ipynb # ML dataset generation and analysis
│   ├── 04_ml_training.ipynb        # Model training and evaluation
│   └── 05_ml_integration.ipynb     # ML-enhanced optimization and visualization
│
├── data/
│   ├── raw/                        # Raw data (warehouse layout, distance matrices)
│   └── processed/                  # Processed data (optimized routes, ML models, results)
│
├── assets/                         # Screenshots and visualizations
│
├── report/                         # Project report
│
├── run_all_notebooks.py            # Script to execute all notebooks in order
├── requirements.txt                # Python dependencies
└── README.md                       # This file

Installation

Prerequisites

• Python 3.10 or higher
• pip package manager

Setup

1. Clone the repository:

   git clone <repository-url>
   cd amazon-robotics-route-optimization-ml

2. Install dependencies:

   pip install -r requirements.txt

3. Verify installation:

   python -c "import ortools, sklearn, numpy, pandas, matplotlib; print('[OK] All dependencies installed')"

Usage

Quick Start

1. Generate warehouse simulation:

   python src/simulate_warehouse.py

   This creates:

   • Warehouse layout with robots, pick-up points, and delivery stations
   • Distance matrix (Euclidean or Manhattan)
   • Visualization saved to data/raw/warehouse_layout.png

2. Run route optimization (baseline):

   python src/optimize_routes.py

   This solves TSP/VRP using OR-Tools and saves:

   • Optimized routes CSV
   • Route visualization
   • Optimization metadata

3. Execute ML pipeline (recommended):

   python run_all_notebooks.py

   This executes all notebooks in order:

   • 03_ml_data_generation.ipynb: Generate ML dataset
   • 04_ml_training.ipynb: Train ML models
   • 05_ml_integration.ipynb: Integrate ML with optimization

   Alternatively, run notebooks individually in Jupyter.

Manual Execution

For step-by-step execution:

1. Generate ML dataset:

   • Open notebooks/03_ml_data_generation.ipynb
   • Run all cells to generate synthetic dataset with congestion, obstacles, and travel times

2. Train ML models:

   • Open notebooks/04_ml_training.ipynb
   • Run all cells to train Random Forest and MLP models
   • Models are saved to data/processed/ml_model_rf.pkl

3. ML-enhanced optimization:

   • Open notebooks/05_ml_integration.ipynb
   • Run all cells to:
     • Build ML-based cost matrix
     • Solve optimization with ML predictions
     • Compare baseline vs ML routes
     • Generate comprehensive visualizations

Results and Performance

ML Model Performance

• Random Forest Regressor:

  • RMSE: ~15-25% of mean travel time
  • MAE: ~10-20% of mean travel time
  • R²: 0.75-0.90

• MLP Regressor:

  • Comparable performance to Random Forest
  • Better generalization on unseen congestion patterns

Optimization Results

• Baseline (Distance-based):

  • Uses simple Euclidean/Manhattan distances
  • Fast computation
  • May not account for congestion

• ML-Enhanced:

  • Uses predicted travel times considering congestion and obstacles
  • Routes adapt to warehouse conditions
  • Typically 5-15% improvement in congested scenarios
  • Equivalent or better performance in normal conditions

Visualizations

The project generates several visualizations:

1. Warehouse Layout: Initial warehouse configuration with all points
2. Route Comparison: Side-by-side comparison of baseline vs ML routes
3. ML Error Analysis: 4-panel plot showing prediction accuracy
4. Congestion Heatmaps: Warehouse congestion distribution with route overlays

All visualizations are saved to data/processed/.

Technical Details

Technologies Used

• Python 3.10+: Core programming language
• OR-Tools: Google's optimization library for TSP/VRP solving
• scikit-learn: Machine learning models (RandomForest, MLP)
• NumPy & Pandas: Data manipulation and numerical computing
• Matplotlib & Seaborn: Data visualization
• Jupyter Notebooks: Interactive analysis and documentation

Key Algorithms

1. TSP/VRP Solving:

   • OR-Tools Routing Solver
   • Search strategies: PATH_CHEAPEST_ARC, GUIDED_LOCAL_SEARCH
   • Time limits: 30 seconds (configurable)

2. ML Models:

   • Random Forest Regressor (100 trees, max_depth=10)
   • MLP Regressor (2 hidden layers, 50 neurons each)
   • Features: coordinates, distance, congestion, obstacles, robot count

3. Distance Metrics:

   • Euclidean distance (default)
   • Manhattan distance (optional)

Project Objectives

This project demonstrates:

1. Operations Research: Application of TSP/VRP algorithms to warehouse logistics
2. Machine Learning: Predictive modeling for travel time estimation
3. Hybrid Systems: Combining traditional optimization with ML predictions
4. Real-world Application: Simulation of Amazon Robotics warehouse scenarios
5. End-to-end Pipeline: Complete workflow from data generation to visualization

Performance Metrics

• Dataset Size: 1,000-10,000 samples (configurable)
• Warehouse Grid: 20×20 (configurable)
• Optimization Time: <30 seconds for typical problems
• ML Training Time: <5 minutes for 10K samples
• Model Accuracy: R² > 0.75, MAE < 20% of mean

Future Improvements

• Real-time congestion updates
• Multi-objective optimization (time, energy, priority)
• Reinforcement learning for adaptive routing
• Integration with actual warehouse data
• Distributed optimization for large-scale warehouses

Acknowledgments

• Google OR-Tools for optimization algorithms
• scikit-learn team for ML tools
• Amazon Robotics for inspiration and real-world context

---

Note: This project simulates warehouse operations and is not connected to actual Amazon systems. All data is synthetic and generated for demonstration purposes.