Reproduce “CorrectionPlanner: Self-Correction Planner with Reinforcement Learning in 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
Yihong Guo, Dongqiangzi Ye, Sijia Chen, Anqi Liu, Xianming Liu
Researchers from Johns Hopkins University and XPENG Motors developed CorrectionPlanner, an autonomous driving system that integrates an explicit self-correction mechanism inspired by LLM reasoning. This system reduced collision rates by over 20% in Waymax reactive simulations and achieved state-of-the-art planning scores on the nuPlan dataset by iteratively refining unsafe motion proposals.
Key insights
- An explicit self-correction mechanism, adapted from LLM reasoning to the motion-token space, can significantly enhance safety by allowing autonomous driving planners to revise unsafe actions pre-execution.
- Leveraging a "correction trace" (a history of rejected unsafe tokens) enables the policy to learn from past failures and shift its subsequent proposals away from unsafe regions more effectively than simple rejection sampling.
- Combining imitation learning with model-based reinforcement learning and a reactive world model is crucial for teaching proactive self-correction and exploring counterfactual unsafe scenarios not present in expert data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.