P-Agent: Training an Autonomous agent to Recognize + Fetch a Package using Pedras and AirSim

Fall Team: Michael Lee, Adam Guo, Vani Sachdev, Christine Cannon Spring Team: Michael Lee, Jared Mejia, Danica Du, Rachel Yang, Nessa Kiani

Project Layout

data: Stores the results of Data Collection
Data_Collection: Scripts that run data collection
Network
-CNN
-Hardcoded (Controller)
-Reinforcement Learning Algorithm (Replaces Hardcoded)
Unreal_Envs: Holds project environments\

How To:

1. Open a packaged unreal environment located in unreal_envs to start the simulation (F1 opens options)
2. Open and execute the python script you wish to run
3. Exit the simulation by typing "~ Exit"

Approach

We use a Convolutional Neural Network to learn images of packages and extract information of the agent's relative pitch, yaw roll. A hardcoded controller executes protocols (e.g. If it doesn't have a direct line of sight, it should rotate (rather than hit a wall)).