Research map / Reinforcement learning / Visual and representation RL
HAD: Combining Hierarchical Diffusion with Metric-Decoupled RL for End-to-End Driving
Wenhao Yao, Xinglong Sun, Zhenxin Li, Shiyi Lan, Zi Wang, Jose M Alvarez, Zuxuan Wu
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 |
|---|---|---|---|
| expand top-K anchors in polar coordinates instead of adding random noise (ablation) | HAD · NAVSIM v2 | EPDMS: 84.9 -> 88.6 | Improved |
| expand in polar rather than Cartesian (XY) coordinates (ablation) | HAD · NAVSIM v2 | EPDMS: 85.9 -> 88.6 | Improved |
| restrict first-stage candidates to K=2 instead of all 20 (ablation) | HAD · NAVSIM v2 | EPDMS: 79.8 -> 88.6 | Improved |
| increase inference expansion density from 5x5 to 7x7 (ablation) | HAD · NAVSIM v2 | EPDMS: +0.1 EPDMS | No clear effect |
| decouple RL reward across metrics (MDPO) instead of coupled PDMS reward (ablation) | HAD · NAVSIM v2 | EPDMS: 87.8 -> 88.6 | Improved |
| use distance-only reward for RL (ablation) | HAD · NAVSIM v2 | EPDMS: 87.8 -> 62.6 (vs PDMS reward) | Hurt |
| add distance to PDMS in a coupled reward (ablation) | HAD · NAVSIM v2 | EPDMS: 87.8 -> 86.9 | Hurt |
| add safety metric scoring to refinement stage (ablation) | HAD · NAVSIM v2 | EPDMS: 86.7 -> 87.3 | Improved |
| add RL training in the local refinement stage (ablation) | HAD · NAVSIM v2 | EPDMS: 87.3 -> 88.6 | Improved |
| add safety and RL supervision to the first (intention) stage (ablation) | HAD · NAVSIM v2 | EPDMS: 88.6 -> 87.2 | Hurt |
| replace online simulator rewards with offline reward retrieval (ablation) | HAD · training efficiency | training time (h): 64.4 -> 13.6 | Improved |
| drop LiDAR input (camera-only HAD-L) (ablation) | HAD · NAVSIM v1 | PDMS: 90.2 -> 89.9 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- ZTRS — ZTRS: Zero-Human Demonstration End-to-end Autonomous Driving with Trajectory Scorer (ECCV 2026) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Diffusion-ES: Gradient-free Planning with Diffusion for Autonomous and Instruction-guided Driving (CVPR 2024 · similarity 0.91)
- DiffusionDrive: Truncated Diffusion Model for End-to-End Autonomous Driving (CVPR 2025 · similarity 0.90)
- RAD: Training an End-to-End Driving Policy via Large-Scale 3DGS-based Reinforcement Learning (NeurIPS 2025 · similarity 0.90)
- ChainFlow-VLA: Causal Flow Planning with Vision-Language Models (NeurIPS 2026 · similarity 0.90)
- Diffusion-Based Planning for Autonomous Driving with Flexible Guidance (ICLR 2025 · similarity 0.89)
- CarPlanner: Consistent Auto-regressive Trajectory Planning for Large-Scale Reinforcement Learning in Autonomous Driving (CVPR 2025 · 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.
