InsightDrive: Insight Scene Representation for End-to-End Autonomous Driving
Ruiqi Song*, Xianda Guo*, Yanlun Peng, Hangbin Wu
$\dagger$ , Qinggong Wei, Long Chen$\dagger$
* Equal contribution
- [2025/5/07] README.md Release
![overview]
- We present InsightDrive, which leverages CoT instructions to fine-tune LLMs and establishes a human–LLM–vehicle distillation pipeline that transfers human driving cognition into onboard models for joint explicit and implicit scene representation.
- We design a Task-level Mixture-of-Experts adapter that injects human driving cognitive processes into scene representations with minimal parameter overhead, which enhances scene understanding and reasoning.
- We propose a diffusion-based generative planner that uses explicit attention and implicit reasoning as conditions for generating robust and adaptive trajectories.
- We conduct comprehensive experiments on both the nuScenes and Navsim benchmarks, which demonstrate the effectiveness and robustness of InsightDrive and show that it achieves state-of-the-art performance.
Also thanks to these excellent open-sourced repos: GenAD
If you find this project helpful, please consider citing the following paper:
@article{insightdrive2025,
title={InsightDrive: Insight Scene Representation for End-to-End Autonomous Driving},
author={Ruiqi Song and Xianda Guo and Hangbin Wu and Qinggong Wei and Long Chen },
journal={https://arxiv.org/abs/2503.13047},
year={2025}
}