Collision Avoidance with AirSim

Reference: https://github.com/simondlevy/AirSimTensorFlow

This repository demonstrates how to use Microsoft AirSim to collect image data from a simulated vehicle, train a neural network using TensorFlow, and perform real-time collision detection and multi-car simulations.

---

🧰 Prerequisites

• Hardware: Recommended hardware for Unreal Engine 4 is required for running AirSim. We recommend using the precompiled AirSimNH Windows binary.
• Python: Use Python 3.8 (64-bit) to ensure compatibility with TensorFlow and other dependencies.
• Simulator: AirSim, launched via run.bat in the AirSimNH folder.

---

⚙️ Setting Up the Conda Environment

To isolate dependencies and ensure compatibility, create and activate a Conda environment named games:

conda create -n games python=3.8
conda activate games

---

🧪 Installing Dependencies

After activating the environment, install all required packages using the provided requirements.txt:

pip install -r requirements.txt

This will install known-compatible versions of:

• tensorflow==2.6.2
• numpy==1.19.2
• six==1.15.0
• matplotlib==3.3.4
• airsim==1.8.1
• opencv-contrib-python==4.5.3.56
• and other necessary dependencies

---

🚀 Instructions for Running the Project

1. image_collection.py

• Launch AirSimNH using run.bat
• Run this script to start collecting first-person-view images
• The car drives forward and saves images to the carpix directory
• Stops when a collision is detected

2. collision_training.py

• Converts images to grayscale
• Labels them as “safe” or “collision”
• Trains a softmax neural network with TensorFlow
• Saves the model parameters as params.pkl

3. collision_testing.py

• Loads params.pkl
• Reconstructs the trained network
• Feeds live images from AirSim
• Stops the vehicle before collision is predicted

4. CollisionAvoidance.py

• Creates a second car that backs out of a parking space
• Runs in parallel with the main car using threads
• Includes potential for further autonomous behaviors

---

🧠 How It Works

• Images collected before a crash are labeled "safe", and the final image is labeled "collision"
• A softmax neural net is trained using TensorFlow 1.x API via tensorflow.compat.v1
• The trained model is saved using pickle
• During testing, live images are passed through the model to predict collisions and apply brakes
• CollisionAvoidance.py demonstrates multi-car behavior using AirSim APIs

---

🧊 Freezing Your Environment (Optional)

After setup, you can freeze your environment to recreate it later:

pip freeze > requirements_exact.txt

Then restore with:

pip install -r requirements_exact.txt

---

Happy simulating! 🏎️