Hydra-MDP：基于多目标水蛇蒸馏的端到端多模态规划
计算机视觉与模式识别
2024-09-02 v4
摘要
我们提出Hydra-MDP，这是一种新范式，采用多个教师进行教师-学生模型。该方法使用来自人类和基于规则的教师的知识蒸馏来训练学生模型，该模型具有多头解码器，以学习针对各种评估指标定制的多样化轨迹候选。凭借基于规则教师的知识，Hydra-MDP能够以端到端方式学习环境如何影响规划，而不必依赖非可微的后处理。这一方法在Navsim挑战中取得第1名，显示出在多样化驾驶环境和条件下的显著泛化能力提升。更多细节请访问\url{https://github.com/NVlabs/Hydra-MDP}。
引用
@article{arxiv.2406.06978,
title = {Hydra-MDP: End-to-end Multimodal Planning with Multi-target Hydra-Distillation},
author = {Zhenxin Li and Kailin Li and Shihao Wang and Shiyi Lan and Zhiding Yu and Yishen Ji and Zhiqi Li and Ziyue Zhu and Jan Kautz and Zuxuan Wu and Yu-Gang Jiang and Jose M. Alvarez},
journal= {arXiv preprint arXiv:2406.06978},
year = {2024}
}
备注
The 1st place solution of End-to-end Driving at Scale at the CVPR 2024 Autonomous Grand Challenge