Project
| # | Title | Team Members | TA | Documents | Sponsor |
|---|---|---|---|---|---|
| 31 | Virtually Trained Self-Balancing Pole System |
Henry Thompson Kishora Adimulam Mason Ryan |
Amr Martini | design_document2.pdf final_paper3.pdf other2.pdf other1.pdf proposal1.pdf video |
|
| Teammates: Kishora Adimulam (adimula2), Henry Thompson (hcthomp2), Mason Ryan (mjryan5) ##Background: There is a growing use for virtual reality as a training environment for AI for applications in the real world. Game engines like Unity have even released machine learning tool-kits for researchers and developers to experiment training AI inside games and simulations. There has been past work in translating these simulation-trained models to physical systems, such as the project done by OpenAI which taught a robot to stack different colored blocks in a specific order only after seeing it once in a virtual simulation. However, there has been no past projects translating AI models trained in Unity to real physical systems. ##Solution: Our project would be to create a self-balancing pole system, which would be trained as a simulation in Unity and uploaded into a physical system which then would gain the ability to balance the pole. Specifically, we would create a 3D simulation replicating all of the attributes of the physical system itself, and train the agent to learn to balance the pole using the Python API and Tensorflow. The trained Tensorflow model would then be uploaded into a Raspberry Pi, which would then use the control signals of the agent to operate motors to balance the real physical pole. ##Technical Overview: There will be four major subsystems inside the entire project: 1. Unity Engine Simulation/Training The Unity engine will train a 3D simulation of a replica of our hardware system. The design of the system will be a one-dimensional track, with a motor that can travel in either direction as well as a pole will be attached to the motor by a cylindrical joint. The system will be trained to swing up a pole which is initially hanging down from the joint, and calibrate itself to keep the pole in a vertical position. We will be using the Proximal Policy Optimization algorithm to train the agent, since it has been shown by OpenAI to be one of the best performing and easy to tune reinforcement learning algorithms. The resulting trained Tensorflow model will be used as the input into the Raspberry Pi Interface. 2. Unity to Raspberry Pi Interface This interface will take the trained Tensorflow model output by the Unity Engine to create a mapping between the state (angle, angular velocity, and position) to the output (voltage applied to motor). The Raspberry Pi will store this mapping so it can communicate with the hardware system. 3. Raspberry Pi to Hardware System Interface The hardware system is going to need to communicate the position of the cart, and the angle and angular velocity of the pole so that the Raspberry Pi can determine the output (voltage applied to the motor) to keep the pole balanced. The Raspberry Pi will output a voltage to the to the motor on the cart to set its velocity. 4. Hardware System The hardware system will be a replica of our 3D simulated model, involving a movable cart on one-dimensional platform and a pole connected to the motor by a cylindrical joint. It will take inputs from the Raspberry Pi to move the cart in order to balance the pole. The hardware system will need to be able to send its current state back to the Raspberry Pi so that it can determine the its velocity. The system will also contain a gyro sensor that will determines its angular position and velocity and relay it to the controller. Link to Original Post: https://courses.engr.illinois.edu/ece445/pace/view-topic.asp?id=31235 |
|||||