儘管Openpilot在現實世界中成功部署並支援超過250種汽車型號，但它在某些情況下面臨挑戰，如準確檢測自行車和行人、在彎曲道路上執行高速機動以及應對突然變道。此外，它在夜間條件下難以檢測到前車。儘管Openpilot是一個開源項目，但他們仍然保持其訓練代碼私有，並僅公開了他們使用的數據集的一小部分，即Comma2K19。 OP-Deepdive展示了一種使用端到端方法訓練類似Openpilot的模型的方法。
然而，由於數據集的限制，他們從頭開始訓練模型並未包含前車預測。為了應對這些挑戰，我們提出了一種高效的方法來訓練Openpilot模型，Supercombo。我們提出了一個新穎的適配器模塊來微調Supercombo，以保留Openpilot使用的全面真實世界數據集訓練的高效骨幹。此外，我們還引入了一個從Carla模擬器中收集的新數據集，覆蓋各種行駛情境和天氣條件。實驗結果表明，我們的方法在軌跡預測和前車預測方面實現了頂尖的性能，而計算資源（FLOPs）僅略微增加了1.9%

Despite its successful real-world deployment and support for over 250 car
models, Openpilot faces challenges in certain conditions, such as accurately detecting bicycles and pedestrians, executing high-speed maneuvers on curved roads,
and responding to sudden vehicle cut-ins. Additionally, it struggles to detect leading vehicles during nighttime conditions. Although Openpilot is an open-source
project, they still keep their training code private and only publish a small portion of the dataset they use, namely Comma2K19. OP-Deepdive has showcased
an approach to training models similar to Openpilot using an end-to-end method.
However, they trained the model from scratch and did not incorporate lead prediction due to dataset limitations.
To address these challenges, we propose an efficient approach to train the
Openpilot model, Supercombo. We propose a novel adapter module for finetuning Supercombo, preserving the well-performing backbone trained on the comprehensive real-world dataset collected by Openpilot. Furthermore, we introduce
a new dataset from the Carla simulator using Carla-Roach, covering diverse driving scenarios and weather conditions. Experimental results demonstrate that
our approach achieves state-of-the-art performance in trajectory prediction and
lead prediction, with a minimal increase of only 1.9% in computational resources
(FLOPs).

[1] H. Chen, R. Tao, H. Zhang, Y. Wang, W. Ye, J. Wang, G. Hu, and M. Savvides, “Conv-adapter: Exploring parameter efficient transfer learning for
convnets,” arXiv preprint arXiv:2208.07463, 2022.
[2] Tesla, “Tesla Autopilot.” https://www.tesla.com/. Accessed: 7 1, 2023.
[3] Volvo, “Volvo.” https://www.volvoautonomoussolutions.com/. Accessed: 7 1, 2023.
[4] Waymo, “Waymo.” https://waymo.com/. Accessed: 7 1, 2023.
[5] Cruise, “Cruise.” https://getcruise.com/. Accessed: 7 1, 2023.
[6] Comma AI, “Openpilot.” https://github.com/commaai/openpilot, 2017.
Accessed: 7 1, 2023.
[7] L. Chen, T. Tang, Z. Cai, Y. Li, P. Wu, H. Li, J. Shi, J. Yan, and Y. Qiao,
“Level 2 autonomous driving on a single device: Diving into the devils of
openpilot,” arXiv preprint arXiv:2206.08176, 2022.
[8] N. Houlsby, A. Giurgiu, S. Jastrzebski, B. Morrone, Q. De Laroussilhe,
A. Gesmundo, M. Attariyan, and S. Gelly, “Parameter-efficient transfer learning for nlp,” in International Conference on Machine Learning,
pp. 2790–2799, PMLR, 2019.
[9] E. J. Hu, yelong shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang,
and W. Chen, “LoRA: Low-rank adaptation of large language models,” in
International Conference on Learning Representations, 2022.
[10] S. Jie and Z.-H. Deng, “Convolutional bypasses are better vision transformer
adapters,” arXiv preprint arXiv:2207.07039, 2022.
[11] Z. Liu, H. Mao, C.-Y. Wu, C. Feichtenhofer, T. Darrell, and S. Xie, “A convnet for the 2020s,” in Proceedings of the IEEE/CVF conference on computer
vision and pattern recognition, pp. 11976–11986, 2022.
[12] J. Ren, M. Zhang, C. Yu, and Z. Liu, “Balanced mse for imbalanced visual regression,” in Proceedings of the IEEE/CVF Conference on Computer
Vision and Pattern Recognition, pp. 7926–7935, 2022.