Where am I now? Since my last update, the hardware of Ironhide is completely finished. I initially encountered problems with the DC actuator not working or being connected, but I was able to troubleshoot and find out that my problem was the lack of a connection between the Raspberry Pi and the standby port of the Pololu Motor Driver. Once I added the connection, I found that power had to be supplied to the standby connection, as well as the PWM module must be activated in order to have the motor working correctly. My final circuit design, and a full picture of Ironhide can be seen below.
Test files for each component of the car were created and have been uploaded to the Ironhide GitHub repository.
Now that the hardware components have been squared away, I put my focus on the network architecture for stop sign detection. The neural network architecture I chose is comprised of 3 Convolutional layers, 3 Convolutional Transpose layers, dropout, maxpooling, and Rectified Linear Units for a baseline activation function. The input to the network is a 256x256 RGB image taken from the on-board camera, and passed through the network, with about 500k trainable parameters. A visualization of the network is shown here.
The lack of accurate training data for my model was the largest roadblock I faced on the software side. Early on in hardware construction I cut the connection between the remote control transmitter and the DC actuators and servo motor. My most recent update is that I was able to implement a keyboard controller that can control Ironhide through SSH.
I also will be utilizing a Proportion-Integral-Derivative controller to keep the car on the constructed track. The PID controller will provide streering and throttle values to Ironhide to prevent it from going off course
What’s Next? I’m loading data with the correct format into my model is where I am currently stuck. My next step will be collecting training data using the keyboard controller I implemented. After training and testing, I want to implement more object detection for chair legs and walls, then a full running loop that integrates the CNN and PID controller to let the car drive itself through the loop.
More importantly, I want to document my entire process from start until now into a single guide called the Ironhide Guide. The purpose of this guide will be as a low-stress, low-cost, introduction for students that want a beginning understanding of robotics to hopefully act as a gateway to a deeper love for robotics engineering.