EUVIMEDEUVIMED entdecken
Crossref

MFPAD: MemoryForgetting Planning for Long-Horizon End-to-End Autonomous Driving

Yikai Wu, Qizhou Hu, Aiguo Lei, Ziying Song

IEEE Latin America Transactions · 2026-08 · 24 (8) · 753-764

DOI: 10.1109/tla.2026.11577664 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Latin America Transactions, 2026-08. Verfasst von Yikai Wu und 3 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Recent planning-oriented end-to-end autonomous driving methods have achieved competitive performance on short-horizon (3 s) trajectory prediction. However, their accuracy and stability degrade markedly when the prediction horizon is extended to long-horizon (6 s), leading to drifting trajectories. A key reason is that most existing frameworks adopt simple feed-forward multilayer perceptron regressors in the trajectory refinement head, which lack explicit modeling of long-term temporal dependencies and planning inertia. To address this limitation, we propose the MemoryForgetting Planning for Autonomous Driving, a plug-in refinement head that combines an LSTM-based memory network with a Transformer-based forgetting network. The memory network autoregressively rolls out a coarse long-horizon trajectory and exposes a sequence of hidden states, while the forgetting network attends over these states and surrounding-agent features with token-level dropout to suppress outdated or noisy motion cues. A lightweight gating module fuses coarse and corrected trajectories at each time step, yielding temporally consistent, interaction-aware plans over a 6 s horizon. We evaluate our method on NuScenes, Adv-NuScenes, Bench2Drive, and NAVSIM, and the results demonstrate consistent improvements. Compared with baselines, it reduces the average collision rate on nuScenes by 11.1% and the 3 s collision rate on Adv-NuScenes by 29.2%, while achieving a 6 s collision rate of 2.01% on NuScenes. In addition, the closed-loop results on Bench2Drive and NAVSIM show that the proposed refinement head also improves downstream driving performance under feedback-driven evaluation. The source code is available at https://github.com/Y1Ka1/MFPAD

Zitieren

APA

Wu, Y., Hu, Q., Lei, A., & Song, Z. (2026). MFPAD: MemoryForgetting Planning for Long-Horizon End-to-End Autonomous Driving. IEEE Latin America Transactions, 24(8), 753-764. https://doi.org/10.1109/tla.2026.11577664

Vancouver

Wu Y, Hu Q, Lei A, Song Z. MFPAD: MemoryForgetting Planning for Long-Horizon End-to-End Autonomous Driving. IEEE Latin America Transactions. 2026;24(8):753-764. doi:10.1109/tla.2026.11577664

BibTeX
@article{Wu2026,
  author = {Wu, Yikai and Hu, Qizhou and Lei, Aiguo and Song, Ziying},
  title = {MFPAD: MemoryForgetting Planning for Long-Horizon End-to-End Autonomous Driving},
  journal = {IEEE Latin America Transactions},
  year = {2026},
  volume = {24},
  number = {8},
  pages = {753--764},
  doi = {10.1109/tla.2026.11577664}
}
Kostenlos starten →