Reproduce “AdaDrive: Self-Adaptive Slow-Fast System for Language-Grounded 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
Ruifei Zhang, Junlin Xie, Wei Zhang, Weikai Chen, Xiao Tan, Xiang Wan, Guanbin Li
AdaDrive from CUHK Shenzhen, Sun Yat-sen University, and Baidu Inc. introduces a self-adaptive slow-fast system for language-grounded autonomous driving that dynamically controls when and how much a Large Language Model (LLM) contributes to planning. This approach achieves superior driving performance on the LangAuto benchmark while significantly reducing memory usage by up to 75% and inference time by up to 64% compared to prior LLM-enhanced methods.
Key insights
- Dynamically activating LLMs only when contextually necessary, guided by an adaptive activation loss, significantly improves efficiency without sacrificing performance.
- Modulating LLM influence continuously, rather than a binary on/off, enables more nuanced and effective integration with conventional planners.
- Specialized visual encoders like LS-Qformer and memory management strategies like PMF are critical for efficient real-time processing of streaming data in LLM-enhanced autonomous driving.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.