Core Concepts
A novel game-based end-to-end learning and testing framework for autonomous vehicle highway driving, by learning from human driving skills.
Abstract
The researchers propose a game-based end-to-end learning and testing framework for autonomous vehicle highway driving, using the popular game Grand Theft Auto V (GTA V).
Key highlights:
They utilize GTA V to collect highway driving data with programmable labels.
An end-to-end architecture is trained to predict steering and throttle values from game screen images.
The predicted control values are sent to the game via a virtual controller to keep the vehicle in lane and avoid collisions.
The proposed framework is validated in GTA V, demonstrating the effectiveness of the game-based approach for learning human driving skills.
The researchers collected a dataset of GTA highway driving with programmable labels, which is shared with the community.
Two neural network architectures, Nvidia's end-to-end network and VGG-19, were trained and tested on the collected data.
The VGG-19 model, leveraging transfer learning, outperformed the Nvidia's architecture in terms of validation loss and inference performance.
The game-based framework allows efficient learning and testing of autonomous driving algorithms without the need for real-world data and infrastructure.
Stats
"Americans as a collective cover nearly 3.2 trillion miles every year and nearly 23% of the driving is done on the highways."
"Americans spend a total of 71 billion hours every year travelling on the highways."
Quotes
"Highways are also suitable for the implementation of autonomous vehicles due to their well-structured roadways and pedestrian free environment."
"Building a specific facility is very costly. Apart from this, significant amount of data is required to train these algorithms which would again lean to higher cost and time."