The paper addresses the limitations of existing autonomous driving research, which has predominantly relied on online reinforcement learning and synthetic datasets. To overcome these limitations, the authors introduce 19 datasets, including real-world human driver datasets from the US Highway 101 (NGSIM) project, and seven popular offline reinforcement learning algorithms in three realistic driving scenarios: highway, lane reduction, and cut-in traffic.
The authors first analyze the NGSIM dataset to extract relevant attributes and properties, such as vehicle lengths, target velocities, and the number of vehicles. They then pre-process the raw data to fit the proposed POMDP model, which includes state, observation, action, reward, and a unified decision-making process that can be applied across different scenarios.
The paper then benchmarks the performance of various offline reinforcement learning algorithms, including Behavioral Cloning (BC), Imitative Learning, Batch Constrained Q (BCQ), Conservative Q Learning (CQL), Implicit Q Learning (IQL), Ensemble-Diversified Actor-Critic (EDAC), and Policy in the Latent Action Space (PLAS), on the provided datasets and driving scenarios. The results offer insights into the usability of real-world datasets in offline reinforcement learning and the performance of different algorithms across various driving conditions.
The authors conclude that the provided datasets and benchmarks can serve as a comprehensive framework for further research in the field of offline reinforcement learning for autonomous driving.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Dongsu Lee,C... at arxiv.org 04-04-2024
https://arxiv.org/pdf/2404.02429.pdfDeeper Inquiries