Reproduce “The Era of End-to-End Autonomy: Transitioning from Rule-Based Driving to Large Driving Models” 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
Eduardo Nebot, Julie Stephany Berrio Perez
This report examines the architectural transition in autonomous driving from traditional rule-based, modular systems to end-to-end learning systems and Large Driving Models (LDMs). It highlights how this shift, driven by the limitations of older paradigms, enables more robust and scalable autonomy, leading to the commercial deployment of supervised E2E driving (L2++) and influencing broader embodied AI applications.
Key insights
- E2E Large Driving Models (LDMs) effectively leverage fleet-scale data to acquire broad driving competence, addressing the long-tail problem and exhibiting emergent behaviors not explicitly programmed.
- The two-step training methodology, combining imitation learning with reinforcement learning for edge cases, enables E2E systems to achieve both natural driving behavior and potentially safer-than-human performance.
- Supervised E2E autonomy (L2++) is emerging as a critical product category, poised for large-scale consumer deployment and demonstrating a credible path to robust performance with reduced hardware costs, particularly with camera-only approaches.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.