Reproduce “DriveMamba: Task-Centric Scalable State Space Model for Efficient 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
Haisheng Su, Wei Wu, Feixiang Song, Junjie Zhang, Zhenjie Yang, Junchi Yan
DRIVEMAMBA presents a task-centric scalable state space model for efficient end-to-end autonomous driving, integrating perception, prediction, and planning into a unified Mamba decoder. The framework achieves competitive performance on nuScenes and Bench2Drive datasets while demonstrating efficiency gains, including being 3.2 times faster and using 68.8% less GPU memory compared to Transformer-based methods.
Key insights
- State Space Models (SSMs), particularly Mamba's linear-complexity operator, provide an efficient and scalable alternative to Transformer attention for handling long spatiotemporal sequences in E2E-AD.
- A task-centric, fully sparse representation, combined with a hybrid spatiotemporal scan and trajectory-guided token ordering, can effectively preserve spatial locality and prioritize planning-relevant interactions.
- Integrating view correspondence learning, dynamic task relation modeling, and long-term temporal fusion within a unified Mamba decoder fosters "planning-oriented perception" and improves overall task synergy in E2E-AD.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.