¶ Self Driving Vehicular Project
Team: Michelle Gutwein, William Zhang, Aashirya Varma, Tanvi Gullapudi, and Aakash Sengupta
Advisors: Ivan Seskar, Jennifer Shane, and Bernhard Firner
Project Description & Goals: The goal of this project is to develop a cost-effective, reduced-scale autonomous vehicle platform that provides a realistic environment for testing and improving self driving algorithms. Using a custom made model car, we trained a convolutional neural network by driving it in the smart city environment. Our focus was on developing a more robust training framework that allows multiple autonomous cars to be trained simultaneously.
Technologies: ROS (Robot Operating System), Pytorch
Documentation:
¶ Week 1 (May 26-28)
- Debugged and completed the Wi-Fi connection and pairing between the car and the training server
- Test ran the car on the previous year's machine learning model
¶ Week 2 (June 1-4)
- Designed the CAD model for the camera holder and 3D printed it
- Created a camera calibration Python script for extrinsic calibration
¶ Week 3 (June 8-11)
- Designed and printed a 3D calibration box to complete the calibration pipeline
- Updated the Parts Key to include all parts required for assembly
¶ Week 4 (June 15-18)
- Debugged software
- Flashed Jetson Orin Nano
- Calibrated camera
¶ Week 5 (June 22-25)
- Recorded data on Rascal
- Built new car, Bandit
- Started migration from ROS 1 to ROS 2
¶ Week 6 (June 29-July 2)
- Started building third car
- Collected more data on Rascal
- Corrected position transform on the camera
- Finalized migration from ROS 2 to create a working clone of the old software
¶ Week 7 (July 6-9)
- Tested Rascal and Bandit
- Created new wallpaper images for the city
- Programmed NPC car to mimic traffic
¶ Week 8 (July 13-16)
- Debugged the evaluator and simulator for testing models
- Finished setting up the third car, Scamp
- Looked into how to track movement and calculate distance on the NPC car
¶ Week 9 (July 20-23)
- Debugged car hardware
- Tested Scamp
- Refined steering calibration
- Set up the Raspberry Pi for the fourth car
- Built ROS 2 setup instructions to support Raspberry Pi (using Docker)
¶ Week 10 (July 27-30)
- Finished building the fourth car, Brat
- Created a setup approach and launch file to support Raspberry Pi camera
- Created an experimental ML validation metric to test intersection detection accuracy
- Created a more accurate augmentation image skewing method
- Debugged training parameters to find optimal model
¶ Final Results
This summer, our main contributions include:
- Created a hardware and software framework to support multiple cars, which will be essential to test advanced dynamic vehicle avoidance in future self-driving progress
- Calibrated each car's camera, turning, and speed to allow each car to run the same machine learning model despite hardware differences
- Set up a new machine learning evaluator that allows faster and easier testing of models in the simulator before running them on the physical road
- Built and redesigned improved hardware for three new cars
- Trained a working model that stops at intersections and performs turns
