Reproduce “CogAD: Cognitive-Hierarchy Guided End-to-End Autonomous Driving” on OpenResearch

About this paper

Zhennan Wang, Jianing Teng, Canqun Xiang, Kangliang Chen, Xing Pan, Lu Deng, Weihao Gu

CogAD, an end-to-end autonomous driving model, emulates human hierarchical cognition for perception and planning, demonstrating superior performance in complex driving scenarios. The model achieves state-of-the-art results in long-tail scenarios, for instance, reducing L2 error by 10% for 3-point turns, and exhibits robust closed-loop driving with a 24.00% success rate on the Bench2Drive benchmark.

Key insights

• Emulating human-like hierarchical cognition, from global scene understanding to local instance details in perception, and from high-level intent to low-level trajectory in planning, significantly enhances E2E-AD robustness.
• Explicitly modeling uncertainty at two distinct levels (intent and trajectory) allows the autonomous system to explore and generate diverse yet plausible multimodal driving options, crucial for complex situations.
• Data-driven intent anchors and shared learnable mode embeddings across planning and motion prediction tasks improve model performance by providing a more generalized and less manually engineered approach to behavioral uncertainty.

Summary by alphaXiv.