Research map / Reinforcement learning / RL from feedback
PlannerRFT: Reinforcing Diffusion Planners through Closed-Loop and Sample-Efficient Fine-Tuning
Hongchen Li, Tianyu Li, Jiazhi Yang, Mingyang Shang, Gaoqiang Wu, Caojun Wang, Haochen Tian, Zengrong Lin, Zhihui Hao, XianPeng Lang, Jia Hu, Hongyang Li
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 |
|---|---|---|---|
| reinforcement fine-tune with policy-guided denoising and GRPO | IL-pretrained Diffusion Planner · Test14-hard (reactive) | nuPlan score: 69.22 -> 72.21 | Improved |
| reinforcement fine-tuning | IL-pretrained Diffusion Planner · Val14 (reactive) | nuPlan score: 82.80 -> 84.46 | Improved |
| reinforcement fine-tuning | IL-pretrained Diffusion Planner · Val14 (non-reactive) | nuPlan score: 89.87 -> 89.96 | No clear effect |
| reinforcement fine-tuning | IL-pretrained Diffusion Planner · Test14-hard (non-reactive) | nuPlan score: 75.99 -> 77.16 | Improved |
| add learned adaptive exploration policy (ablation) | PlannerRFT · Test14-hard (reactive) | R-score: 70.65 -> 72.21 vs fixed Beta; 65.82 vs uniform | Improved |
| add classifier guidance without learned policy (uniform scales) (ablation) | IL-pretrained planner RFT · Test14-hard (reactive) | R-score: 68.83 -> 65.82 | Hurt |
| add guidance with fixed Beta distribution (ablation) | RFT without guidance · Test14-hard (reactive) | R-score: 68.83 -> 70.65 | Improved |
| fine-tune on Lt90 instead of Fail only (ablation) | PlannerRFT · Test14-hard (reactive) | nuPlan score: 63.75 -> 72.21 | Improved |
| fine-tune on Lt90 instead of All scenarios (ablation) | PlannerRFT · Test14-hard (reactive) | nuPlan score: 70.43 -> 72.21 | Improved |
| use RL instead of imitation fine-tuning on Lt90 (ablation) | Diffusion Planner · Test14-hard (reactive) | nuPlan score: 67.55 -> 72.21 | Improved |
| use survival reward instead of terminal reward (ablation) | PlannerRFT · Test14-hard (reactive) | nuPlan score: 71.59 -> 72.21 | Improved |
| lengthen reward horizon from 2 s to 4 s (ablation) | PlannerRFT · Test14-hard (reactive) | nuPlan score: 70.10 -> 72.21 | Improved |
Closest papers by meaning
- Diffusion-ES: Gradient-free Planning with Diffusion for Autonomous and Instruction-guided Driving (CVPR 2024 · similarity 0.93)
- Diffusion-Based Planning for Autonomous Driving with Flexible Guidance (ICLR 2025 · similarity 0.90)
- Task-Agnostic Pre-training and Task-Guided Fine-tuning for Versatile Diffusion Planner (ICML 2025 · similarity 0.90)
- Prior-Guided Diffusion Planning for Offline Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
- ReflectDrive-2: Reinforcement-Learning-Aligned Self-Editing for Discrete Diffusion Driving (NeurIPS 2026 · similarity 0.89)
- DiRecT: Safe Diffusion-Based Planning via Receding-Horizon Denoising (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.
