Research map / Vision-language and multimodal / Robot action and navigation
MindDrive: A Vision-Language-Action Model for Autonomous Driving via Online Reinforcement Learning
haoyu fu, Diankun Zhang, Zongchuang Zhao, Jianfeng Cui, Hongwei Xie, Bing Wang, Guang Chen, Hangjun Ye, Dingkang Liang, Xiang Bai
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 online RL after imitation learning (ablation) | MindDrive (0.5B) · Bench2Drive | Driving Score / Success Rate: 75.85 -> 78.04 DS; 49.30 -> 55.09 SR | Improved |
| add online RL after imitation learning (ablation) | MindDrive-L (3B) · Bench2Drive | Driving Score / Success Rate: 75.78 -> 80.59 DS; 50.71 -> 58.26 SR | Improved |
| add collision penalty (ablation) | MindDrive (0.5B) · Bench2Drive | Success Rate (%): 49.30 -> 50.70 | Improved |
| add traffic light penalty on top of collision (ablation) | MindDrive (0.5B) · Bench2Drive | Success Rate (%): 50.70 -> 51.47 | Improved |
| add route deviation penalty (ablation) | MindDrive (0.5B) · Bench2Drive | Success Rate (%): 51.47 -> 51.85 | Improved |
| add stop sign penalty (ablation) | MindDrive (0.5B) · Bench2Drive | Success Rate (%): 51.85 -> 55.09 | Improved |
| use KL regularization instead of vanilla PPO (ablation) | MindDrive (0.5B) · Bench2Drive | Success Rate (%): 46.73 -> 55.09 | Improved |
| use KL regularization instead of entropy regularization (ablation) | MindDrive (0.5B) · Bench2Drive | Success Rate (%): 49.24 -> 55.09 | Improved |
| use VLM meta-actions instead of navigation commands (IL) (ablation) | MindDrive (0.5B) · Bench2Drive | Success Rate (%): 41.59 -> 49.30 | Improved |
| set GAE lambda to 1.00 (Monte Carlo) instead of 0.90 (ablation) | MindDrive (0.5B) · Bench2Drive | Success Rate (%): 48.38 -> 55.09 | Improved |
| set GAE lambda to 1.00 instead of 0.95 (ablation) | MindDrive (0.5B) · Bench2Drive | Success Rate (%): 49.03 -> 55.09 | Improved |
| train online RL for 2 epochs instead of 1 (ablation) | MindDrive (0.5B) · Bench2Drive | Success Rate (%): 48.61 -> 55.09 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 4 of 4 on most shared results.
- ORION — ORION: A Holistic End-to-End Autonomous Driving Framework by Vision-Language Instructed Action Generation (ICCV 2025) · better on 3 of 3 shared results
- RecogDrive — ReCogDrive: A Reinforced Cognitive Framework for End-to-End Autonomous Driving (ICLR 2026) · better on 2 of 2 shared results
- UniDrive-WM — UniDrive-WM: Unified Understanding, Planning and Generation World Model For Autonomous Driving (ECCV 2026) · better on 2 of 2 shared results
- AutoVLA — AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning (NeurIPS 2025) · 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
- AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning (NeurIPS 2025 · similarity 0.93)
- Drive-R1: Bridging Reasoning and Planning in VLMs for Autonomous Driving with Reinforcement Learning (AAAI 2026 · similarity 0.92)
- LaST-VLA: Thinking in Latent Spatio-Temporal Space for Vision-Language-Action in Autonomous Driving (NeurIPS 2026 · similarity 0.90)
- Distill to Think, Foresee to Act: Cognitive-Physical Reinforcement Learning for Autonomous Driving (NeurIPS 2026 · similarity 0.90)
- $AutoDrive\text{-}P^3$: Unified Chain of Perception–Prediction–Planning Thought via Reinforcement Fine-Tuning (ICLR 2026 · similarity 0.90)
- NoRD: A Data-Efficient Vision-Language-Action Model that Drives without Reasoning (CVPR 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.
