Reproduce “GEMINUS: Dual-aware Global and Scene-Adaptive Mixture-of-Experts for End-to-End Autonomous Driving” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chi Wan, Yixin Cui, Jiatong Du, Shuo Yang, Yulong Bai, Peng Yi, Nan Li, Yanjun Huang
A framework named GEMINUS addresses the "mode averaging" problem in end-to-end autonomous driving by using a dual-aware Mixture-of-Experts architecture. This system achieves competitive closed-loop performance on the Bench2Drive benchmark with monocular vision, enhancing scenario-specific adaptability and providing robust fallback in ambiguous situations.
Key insights
- The dual-aware Mixture-of-Experts architecture successfully mitigates the "mode averaging" problem by enabling scenario-specific policy generation.
- Combining a Global Expert with Scene-Adaptive Experts and an uncertainty-aware router provides a balanced approach for both adaptive and robust behavior in diverse and ambiguous driving situations.
- A specially tailored MoE framework with explicit scenario and uncertainty awareness is effective for end-to-end autonomous driving, overcoming limitations of generic MoE applications.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.