Reproduce “Learning to Drive in a Day” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Alex Kendall, Jeffrey Hawke, David Janz, Przemyslaw Mazur, Daniele Reda, John-Mark Allen, Vinh-Dieu Lam, Alex Bewley and 1 more
This paper presents the application of deep reinforcement learning to a real autonomous vehicle, demonstrating that a car can learn to lane follow from scratch using only monocular camera input and a sparse reward function. The system learned a robust policy on a private road within a day of training, showcasing rapid real-world adaptation.
Key insights
- A sparse, general reward function (distance traveled without intervention) is sufficient for a DRL agent to learn complex real-world driving behaviors.
- Learned state representations, such as those from a Variational Autoencoder (VAE), significantly boost data efficiency for DRL in complex real-world visual environments, enabling faster learning.
- Hyperparameters tuned in simulation can effectively transfer to a real-world vehicle, reducing the need for extensive real-world tuning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.