【论文精读】PARA-Drive: Parallelized Architecture for Real-time Autonomous Driving

背景信息

团队：英伟达
会议：CVPR2024

主要贡献

1. 分析了目前SOTA的网络结构：包括使用模块、模块位置、模块间信息流类型。其中，信息流类型分为【有具体含义的compact outputs】和【无具体含义的latent features】
2. 评估方式：不同论文针对L2 error等指标不一致

评估方式

VAD在时间间隔上做了平均，导致其结果明显好于UniAD；
UniAD在计算碰撞率时，去掉了和行人的碰撞，导致碰撞率好于VAD；
VAD筛选了输入数据，如果某数据中有一帧数据有问题，就不测试了，将其结果赋值为0，导致错误率降低。
因此，评估方式不一致导致无法横向比较，这篇论文统一了评估方式再比，是很有想法的。

模型设计

1. TTO导致锯齿形的轨迹。test-time optimization (TTO) is not in the training process, and it tends to generate trajectories that deviate from human driving logs. In our supplementary material, we present visualizations showing that TTO often results in zigzag-like trajectories near multiple objects, thereby increasing the L2 error and cannot guarantee to avoid collision.
2. 模块必要性。通过消融实验表明，mapping/occ/motion prediction都是对于planning来说必要的。the removal of any auxiliary task could lead to a significant performance drop on the planning task.
3. 信息流。传递latent features比compact outputs更好。passing the latent queries with high dimensional information tends to bring a stronger performance increase compared to using the compact output representations alone as inputs to planning.
4. 模块位置。模块可以并行，也可以串行。Similar state-of-the-art performance achieved in hybrid designs can also be achieved with a parallel design.
5. 用历史轨迹信息可以boost performance。using CAN bus (velocity, acceleration, angular velocity, etc.) and history trajectories can improve planning. 评价：这篇论文依然是开环，所以还不知道用历史轨迹会导致闭环效果下降。