隨著自動駕駛技術的演進，如何在高動態且具不確定性的環境中精準預測介入行為，並確保系統能於運算資源受限的邊緣端設備即時運行，已成為人機共駕領域的重要挑戰。本研究提出一個結合液態時間常數網路（Liquid Time-Constant Networks, LTC）與多任務學習（Multi-task Learning）的輕量化預測架構，旨在達成高效能的自動駕駛介入預警。
本研究利用 CARLA 模擬器構建單車單意圖之邊界情境，採集包含車輛動力學指標、操控指令與環境語義之結構化數據。在模型設計上，透過雙頭（Dual-head）架構同時執行未來「軌跡回歸」與「介入行為分類」兩項任務，使模型在學習過程中被迫理解車輛運動學約束，從而提升分類任務的穩定性與決策的可解釋性。
實驗結果顯示，LTC具備卓越的連續時間動力學建模能力，在僅 32 個神經元單元（Units）的規模下即可達成效能飽和。相比傳統離散時間模型，本架構不僅能更精確地擬合非線性物理動態，且在 CPU 環境下有不錯的能效，完成運算僅需數毫秒。依據本研究的結果，透過適當的多任務權重配置與模型規模優化，LTC 架構能以較低的運算成本實現具備物理穩健性的即時預警，為量產級車載冗餘監控系統提供了未來在自動駕駛危險偵測實務應用上的參考。

As autonomous driving technology evolves, accurately predicting intervention behavior within highly dynamic and uncertain environments—while ensuring real-time execution on resource-constrained edge devices—has emerged as a critical challenge in the field of shared control. This study proposes a lightweight predictive framework that integrates Liquid Time-Constant (LTC) networks with multi-task learning, aiming to achieve high-performance intervention warning for autonomous vehicles.
Utilizing the CARLA simulator, this research constructs single-vehicle, single-intent edge cases to collect structured data encompassing vehicle dynamics metrics, control commands, and environmental semantics. In terms of model design, a dual-head architecture is developed to simultaneously perform future trajectory regression and intervention classification. This framework forces the model to understand vehicle kinematic constraints during the learning process, thereby enhancing the stability of the classification task and the interpretability of its decisions.
Experimental results demonstrate that LTC networks possess exceptional continuous-time dynamics modeling capabilities, reaching a performance saturation point at a scale of only 32 units. Compared to conventional discrete-time models, the proposed framework not only fits non-linear physical dynamics more precisely but also exhibits good efficiency in CPU environments, completing inference within a few milliseconds. The findings of this study suggest that, through proper multi-task weight configuration and model scale optimization, the LTC framework can achieve physically robust, real-time alerts at a lower computational cost, providing a reference for future application in production-grade, in-vehicle redundant monitoring systems for hazard detection in autonomous driving practice.

[1] B. Paden, M. Čáp, S. Z. Yong, D. Yershov, and E. Frazzoli, "A Survey of Motion Planning and Control Techniques for Self-Driving Urban Vehicles," IEEE Transactions on Intelligent Vehicles, vol. 1, no. 1, pp. 33-55, 2016.
[2] F. Codevilla, M. Müller, A. López, V. Koltun, and A. Dosovitskiy, "End-to-End Driving via Conditional Imitation Learning," in Proc. IEEE Int. Conf. Robot. Autom. (ICRA), 2018, pp. 4693-4700.
[3] V. Talpaert, I. Sobh, B. R. Kiran, P. Mannion, S. Yogamani, A. El-Sallab, and P. Perez, "Exploring Applications of Deep Reinforcement Learning for Real-World Autonomous Driving Systems," in Proc. Int. Conf. Comput. Vis. Theory Appl. (VISAPP), 2019, pp. 298–305.
[4] M. Bansal, A. Krizhevsky, and A. Casagrande, "ChauffeurNet: Learning to Drive by Imitating the Best and Synthesizing the Worst," in Proc. Robotics: Science and Systems (RSS), 2018.
[5] J. Gao, C. Sun, H. Zhao, Y. Shen, D. Anguelov, C. Li, C. Schmid , "VectorNet: Encoding HD Maps and Agent Dynamics from Vectorized Representations," in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2020, pp. 11525-11533.
[6] R. Hasani, M. Lechner, A. Amini, D. Rus, and R. Grosu, "Liquid Time-Constant Networks," in Proc. AAAI Conf. Artif. Intell. (AAAI), vol. 35, no. 9, pp. 7657-7666, 2021.