Research map / Vision-language and multimodal / Robot action and navigation
ReAL-AD: Towards Human-Like Reasoning in End-to-End Autonomous Driving
Yuhang Lu, Jiadong Tu, Yuexin Ma, Xinge Zhu
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 Strategic Reasoning Injector (ablation) | VAD-Base + MiniCPM-V · Bench2Drive val | Avg L2 (m): 1.01 -> 0.89 | Improved |
| add Tactical Reasoning Integrator (ablation) | VAD-Base + MiniCPM-V · Bench2Drive val | Avg L2 (m): 1.01 -> 0.97 | Improved |
| add Hierarchical Trajectory Decoder instead of MLP decoder (ablation) | VAD-Base + MiniCPM-V · Bench2Drive val | Avg L2 (m): 1.01 -> 0.93 | Improved |
| add all three modules together (ablation) | VAD-Base + MiniCPM-V · Bench2Drive val | Avg collision rate (%): 0.21 -> 0.12 | Improved |
| add similarity loss with GT trajectory encoder for strategy features (ablation) | SRI module · Bench2Drive val | Avg L2 (m): 0.92 -> 0.89 | Improved |
| use MSE loss instead of cosine similarity loss (ablation) | SRI module · Bench2Drive val | Avg collision rate (%): 0.17 -> 0.20 | Hurt |
| use KL divergence loss instead of cosine similarity loss (ablation) | SRI module · Bench2Drive val | Avg L2 (m): 0.89 -> 0.89 | No clear effect |
| increase decoder layers from two to three (ablation) | Hierarchical Trajectory Decoder · Bench2Drive val | Avg L2 (m): 0.93 -> 0.97 | Hurt |
| increase decoder layers from one to two (ablation) | Hierarchical Trajectory Decoder · Bench2Drive val | Avg L2 (m): 0.99 -> 0.93 | Improved |
| encode tactical commands as probabilities instead of one-hot (ablation) | TRI module · Bench2Drive val | Avg collision rate (%): 0.16 -> 0.20 (L2 0.97 -> 0.86) | Mixed |
| replace MLP adapter with self-attention adapter (ablation) | SRI module · Bench2Drive val | Avg L2 (m): 0.89 -> 0.90 | No clear effect |
| use multi-view input to the VLM instead of front view (ablation) | ReAL-AD (VAD & Qwen-VL) · Bench2Drive val | Avg L2 (m): 0.53 -> 0.51 | No clear effect |
Closest papers by meaning
- Drive-R1: Bridging Reasoning and Planning in VLMs for Autonomous Driving with Reinforcement Learning (AAAI 2026 · similarity 0.91)
- Continuously Learning, Adapting, and Improving: A Dual-Process Approach to Autonomous Driving (NeurIPS 2024 · similarity 0.91)
- AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning (NeurIPS 2025 · similarity 0.91)
- DriveCombo: Benchmarking Compositional Traffic Rule Reasoning in Autonomous Driving (CVPR 2026 · similarity 0.90)
- SGDrive: Scene-to-Goal Hierarchical World Cognition for Autonomous Driving (CVPR 2026 · similarity 0.90)
- nuReasoning: A Reasoning-Centric Dataset and Benchmark for Long-Tail Autonomous Driving (NeurIPS 2026 · similarity 0.90)
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.
