Core Concepts
End-to-end driving dataset DriveCoT and model DriveCoT-Agent enhance interpretability and performance in autonomous driving.
Abstract
The content introduces the DriveCoT dataset, focusing on chain-of-thought reasoning in end-to-end driving. It also presents the DriveCoT-Agent model, emphasizing interpretability and performance improvements. The dataset includes challenging scenarios for evaluation, while the model predicts detailed speed decisions based on various aspects.
Directory:
Abstract
End-to-end driving progress and challenges.
Introduction
Modular vs. end-to-end designs in autonomous driving.
Data Extraction Process
Collection of sensor data, control decisions, and chain-of-thought labels.
Expert Policy Design
Rule-based expert policy for handling challenging scenarios.
Baseline Model Structure
DriveCoT-Agent structure using video inputs for predictions.
Experiment Results
Open-loop evaluation on validation data and closed-loop evaluation on Town05Long benchmark.
Ablation Study Results
Impact of model parameters on performance.
Qualitative Results & Visualization
Performance examples of DriveCoT-Agent in different scenarios.
Stats
ドライブコットデータセットは、1058のシナリオと36Kのラベル付きサンプルを含む。
モデルDriveCoT-Agentは、ビデオ入力を処理し、異なる側面に関する予測を生成する。
Quotes
"We introduce DriveCoT, the first end-to-end driving dataset containing chain-of-thought thinking process labels."
"DriveCoT-Agent exhibits better performance compared to existing methods in both open-loop and closed-loop testing."