Artificial intelligence that can learn to perform robotic tasks from iteracting with the physical world

Brief Overview:

• What is it about robots that make them so robotic?
  In this project, we will teach a robot to learn and perform a simple athletic task using Q-learning. The foundation of this teaching consists of: a reward system, a learning algorithm, and a way for the A.I. to interact with the environment. By providing incentives for the them to do well, we can train the A.I. to become proficient in a variety of tasks. Without this incentive, the A.I. lacks purpose. In the end, the types of behavior that we endorse (and reward) determines the kind of A.I. that is produced.

Requirements:

• Python 3.6.5
• Arduino IDE > 1.8.5

File/Folder Descriptions:

arduino: code uploaded to robot
processing: telemetry and visualization ai_physical.py: perform training on a physical robot
ai_simulated.py: perform training in a simulated environment
core.py: contains core functions like logging, serial communication, etc.
qlearn.py: learning algorithm based on Q-learning
random_search.py: learning algorithm based on random search (aka. Monte Carlo method)

Install Dependencies:

• Python:

pip install gym, matplotlib, pandas, pickle, pyserial, seaborn

• Arduino Libraries: (tutorial)

LSM6 by Pololu version 1.0.0
Balboa32U4 by Pololu version 1.0.1

Installation

git clone https://github.com/khoa-ngo/LearningThroughPlay

Test Simulated Learning

cd ai_pendulum
./ai_simulated.py