Research map / Label-efficient and robust learning / Recognition and long-tail
Beyond Imitation: Learning Safe End-to-End Autonomous Driving from Hard Negatives
Junli Wang, HuaZhihua HuaZhihua, Xueyi Liu, Zebin Xing, Wei Zhang, Kun Ma, Guang Chen, Hangjun Ye, Long Chen, Pengxuan Yang
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 BeyondDrive hard-negative repulsive loss to LTF* (ablation) | LTF* · NAVSIM v1 navtest | PDMS: 88.7 -> 89.7 | Improved |
| add BeyondDrive to DiffusionDrive | DiffusionDrive · NAVSIM v1 navtest | PDMS: 88.1 -> 89.2 | Improved |
| add BeyondDrive to MeanFuser | MeanFuser · NAVSIM v1 navtest | PDMS: 89.0 -> 90.3 | Improved |
| add BeyondDrive to WoTE | WoTE · NAVSIM v1 navtest | PDMS: 88.3 -> 89.2 | Improved |
| add BeyondDrive to DiffusionDrive on reactive benchmark | DiffusionDrive · NAVSIM v2 navtest | EPDMS: 88.3 -> 88.9 | Improved |
| zero-shot transfer: LTFv7 vs LTF on HUGSIM | LTFv7 · HUGSIM overall | HD-Score: 24.8 -> 34.8 | Improved |
| hyperparameter tuning of LTF (cosine schedule with warmup, lower auxiliary weights) (ablation) | LTF · NAVSIM v1 navtest | PDMS: 83.8 -> 87.4 | Improved |
| add first-order-difference trajectory normalization (ablation) | LTF with tuned hyperparameters · NAVSIM v1 navtest | PDMS: 87.4 -> 88.7 | Improved |
| reduce CFG scale from 1.0 to 0.5 when sampling negatives (ablation) | negative generator · valid negatives proportion | Prop_sample: 8.0% -> 78.8% | Improved |
| enlarge sampling noise to 2.0 sigma (ablation) | negative generator · NAVSIM v1 navtest | PDMS: 89.1 -> 89.7; valid negatives 78.8% -> 96.4% | Improved |
| increase RD loss weight from 0 to 5 (imitation weight 10) (ablation) | LTFv7 · NAVSIM v1 navtest | PDMS: 88.7 -> 89.7 | Improved |
| set RD loss weight equal to imitation weight (ablation) | LTFv7 · NAVSIM v1 navtest | PDMS: 89.7 -> 0.1 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- VADv2 — VADv2: End-to-End Autonomous Driving via Probabilistic Planning (ICLR 2026) · better on 1 of 1 shared result
- TransFuser — TransFuser: Imitation With Transformer-Based Sensor Fusion for Autonomous Driving (TPAMI 2022) · better on 1 of 1 shared result
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
- DriveDPO: Policy Learning via Safety DPO For End-to-End Autonomous Driving (NeurIPS 2025 · similarity 0.88)
- AlignDrive: Aligned Lateral-Longitudinal Planning for End-to-End Autonomous Driving (NeurIPS 2026 · similarity 0.88)
- Prioritizing Perception-Guided Self-Supervision: A New Paradigm for Causal Modeling in End-to-End Autonomous Driving (NeurIPS 2025 · similarity 0.87)
- RAD: Training an End-to-End Driving Policy via Large-Scale 3DGS-based Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
- Flow Matching-Based Autonomous Driving Planning with Advanced Interactive Behavior Modeling (NeurIPS 2025 · similarity 0.86)
- Diffusion-ES: Gradient-free Planning with Diffusion for Autonomous and Instruction-guided Driving (CVPR 2024 · similarity 0.86)
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.
