Students at the Massachusetts Institute of Technology are learning about autonomous driving by taking NVIDIA-powered data science workstations for a spin.
In an undergraduate robotics class at MIT, 17 students were organized into three teams and given a miniature racing car. Their task: teach it how to drive by itself through a complex course inside the basement of the university’s Stata Center.
Sertac Karaman, associate professor of aeronautics and astronautics at MIT, wanted to teach students the process of imitation learning, a technique that uses human demonstrations to train a self-driving model.
NVIDIA’s Jetson AGX Xavier and Quadro RTX-powered Data Science Workstation deliver accelerated computing capabilities that allow Karaman and his students to create various AI-powered prototypes.
Students Wheel It in with Data Science Workstations
Through the process of imitation learning, the students needed to teach their car how to autonomously drive by training a TensorFlow neural network. But first, they needed to collect as much data as they could on the indoor course so the cars could learn how to navigate through the hallways and doors of the Stata Center.
Each car was equipped with an NVIDIA Jetson AGX Xavier embedded system-on-a-module for performance-driven autonomous machines. Using a joystick, the students manually drove the small car around the complex course and recorded data through a camera mounted on its frontend.
Then the neural network, based on the NVIDIA PilotNet architecture, processed that data, learning how to map between observation and action — so the car could estimate steering angles based on what its camera sees.
The students used the advanced computing capabilities of the data science workstations, powered by NVIDIA Quadro RTX GPUs, to train their TensorFlow models, which were then deployed on the miniature race cars for on-device AI inference.
The data science workstations provided massive speedups in performance to greatly reduce iteration times. This allowed the students to quickly train and test various models to find the best one for their race car.
“The students were successful in their projects because the time it took for training the models was faster than we’ve ever seen,” said Karaman. “The accelerated computing capabilities of NVIDIA data science workstations allowed the class to iterate multiple times, and the best performing race cars were trained in only a few minutes.”
Karaman plans to teach the robotics class once again this year, using the data science workstations and the pre-installed AI software stack.
Learn more about Quadro RTX-powered data science workstations.