Reproduce “RoG-DAgger: Rollout-Guided Post-Training for End-to-End 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
Liangyu Zhong, Joachim Sicking, Fabian Hueger, Hanno Gottschalk
RoG-DAgger is a rollout-guided post-training framework that enhances end-to-end driving policy robustness by refining DAgger with an expanded expert solution space, solvability-aware takeover triggers, and field-of-view-aligned planning, leading to improved performance across diverse driving benchmarks.
Key insights
- The quality and relevance of expert supervision, considering its feasibility, timeliness, and compatibility with the student's observation space, are critical for effective imitation learning in post-training.
- Proactive failure prevention, enabled by solvability-aware interventions near the "point of no return" with feasible corrective plans, offers a more valuable learning signal than reactive recovery from already failed states.
- Expanding the expert's available action space, particularly for geometric maneuvers beyond longitudinal control, significantly contributes to preventing failures that single-dimensional responses cannot resolve.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.