Research map / Label-efficient and robust learning / Semi-supervised learning
SelfD: Self-Learning Large-Scale Driving Policies From the Web
Jimuyang Zhang, Ruizhao Zhu, Eshed Ohn-Bar
Check 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 |
|---|---|---|---|
| Replace image-plane waypoint prediction and fixed perspective transform with direct BEV planning | CIL planner · nuScenes Singapore | ADE: 1.86 -> 1.14 | Improved |
| Add speed input to the BEV planner (ablation) | BEV Planner · nuScenes Singapore | ADE: 4.64 -> 1.14 | Improved |
| Add command input to a speed-conditioned BEV planner (ablation) | BEV Planner · nuScenes Singapore | ADE: 1.33 -> 1.14 | Improved |
| Add What If pseudo-label sampling and fine-tune on clean labeled data (ablation) | SelfD · nuScenes Singapore | ADE: 1.14 -> 1.00 | Improved |
| Replace What If uniform sampling with visual odometry estimates (ablation) | SelfD · nuScenes Singapore | ADE: 1.00 -> 1.09 | Hurt |
| Replace uniform sampling with prior-based speed and command sampling (ablation) | SelfD · nuScenes Singapore | ADE: 1.00 -> 1.09 | Hurt |
| Fine-tune the YouTube-pretrained policy on clean labeled data (ablation) | SelfD (What If-Uniform) · Waymo | ADE: 2.24 -> 1.65 | Improved |
| Add YouTube self-training to the BEV planner | SelfD · nuScenes Singapore night | ADE: 1.67 -> 1.28 | Improved |
| Add YouTube self-training to the BEV planner | SelfD · Waymo rainy | ADE: 0.70 -> 0.67 | No clear effect |
| Self-train with pseudo-labeled data from Town 1 | BEV Planner · CARLA Town 2 | Success Rate (%): 12 -> 25 | Improved |
| Add unlabeled Town 2 data to self-training (ablation) | SelfD (Town 1) · CARLA Town 2 | Success Rate (%): 25 -> 26 | Improved |
| Repeat pseudo-labeling and fine-tuning for an additional iteration (ablation) | SelfD · nuScenes Singapore | ADE: 1.00 -> 1.07 | Hurt |
Closest papers by meaning
- Learning to Drive is a Free Gift: Large-Scale Label-Free Autonomy Pretraining from Unposed In-The-Wild Videos (CVPR 2026 · similarity 0.90)
- Prioritizing Perception-Guided Self-Supervision: A New Paradigm for Causal Modeling in End-to-End Autonomous Driving (NeurIPS 2025 · similarity 0.90)
- Uncertainty-Guided Never-Ending Learning to Drive (CVPR 2024 · similarity 0.89)
- Enhancing End-to-End Autonomous Driving with Latent World Model (ICLR 2025 · similarity 0.89)
- S4-Driver: Scalable Self-Supervised Driving Multimodal Large Language Model with Spatio-Temporal Visual Representation (CVPR 2025 · similarity 0.88)
- Continuously Learning, Adapting, and Improving: A Dual-Process Approach to Autonomous Driving (NeurIPS 2024 · 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.
