Research map / Reinforcement learning / Visual and representation RL
Reinforced Refinement With Self-Aware Expansion for End-to-End Autonomous Driving
Haochen Liu, Tianyu Li, Haohan Yang, Li Chen, Caojun Wang, Ke Guo, Haochen Tian, H. Y. Li, Hongyang Li, Chen Lv
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add GRPO cost constraints to IL (ablation) | Hydra-MDP refinement · NAVSIM navtest | PDMS: 87.0 -> 89.3 (+2.3) | Improved |
| RL refinement without IL objective (ablation) | Hydra-MDP refinement · NAVSIM navtest | PDMS: 87.0 -> 87.9 (+0.9) | No clear effect |
| add IL objective to GRPO cost and RL (ablation) | R2SE-MDP · NAVSIM navtest | PDMS: 87.9 -> 89.7 (+1.8) | Improved |
| add self-aware expansion mechanism (ablation) | R2SE-MDP · NAVSIM navtest | PDMS: 89.7 -> 91.6 (+1.9) | Improved |
| loosen hard case threshold to epsilon=10 (ablation) | R2SE-MDP · NAVSIM navtest | PDMS: 91.6 -> 90.9 (-0.7) | Hurt |
| train with only 500 hard cases (epsilon 0.45, PDMS=0) (ablation) | R2SE-MDP · NAVSIM navtest | PDMS: 91.2 (+4.7) | Improved |
| replace GPD tail model with power law (ablation) | R2SE expansion · NAVSIM navtest | PDMS: -0.5 PDMS, -0.6 EP | Hurt |
| replace GPD tail model with KDE (ablation) | R2SE expansion · NAVSIM navtest | PDMS: -2.0 PDMS | Hurt |
| replace GPD tail model with percentile (ablation) | R2SE expansion · NAVSIM navtest | PDMS: -5.0 EP and -5.7 PDMS | Hurt |
| set expansion confidence sigma to 0.05 instead of 0.75 (ablation) | R2SE expansion · NAVSIM navtest | PDMS: -1.5 | Hurt |
| set expansion confidence sigma to 0.95 instead of 0.75 (ablation) | R2SE expansion · NAVSIM navtest | PDMS: -1.3 | Hurt |
| add R2SE to VAD on hardest 20% Bench2Drive scenarios | VAD · Bench2Drive hard scenarios | DS: +21.31 DS, +29.22 route completion | Improved |
Closest papers by meaning
- RAD: Training an End-to-End Driving Policy via Large-Scale 3DGS-based Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
- Prioritizing Perception-Guided Self-Supervision: A New Paradigm for Causal Modeling in End-to-End Autonomous Driving (NeurIPS 2025 · similarity 0.88)
- Model-Based Policy Adaptation for Closed-Loop End-to-end Autonomous Driving (NeurIPS 2025 · similarity 0.88)
- Raw2Drive: Reinforcement Learning with Aligned World Models for End-to-End Autonomous Driving (in CARLA v2) (NeurIPS 2025 · similarity 0.88)
- Distill to Think, Foresee to Act: Cognitive-Physical Reinforcement Learning for Autonomous Driving (NeurIPS 2026 · similarity 0.88)
- Beyond Self-Play and Scale: A Behavior Benchmark for Generalization in Autonomous Driving (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
