Research map / Vision-language and multimodal / Robot action and navigation
DriveMoE: Mixture-of-Experts for Vision-Language-Action Model in End-to-End Autonomous Driving
Zhenjie Yang, Yilin Chai, Xiaosong Jia, Qifeng Li, Yuqian Shao, Xuekai Zhu, Haisheng Su, Junchi Yan
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 dynamic camera view selection to Drive-π0 (ablation) | Drive-π0 · Bench2Drive (closed-loop) | Driving Score: 55.85 -> 69.71 | Improved |
| add supervision to the vision router (ablation) | DriveMoE Vision MoE · Bench2Drive (closed-loop) | Driving Score: 69.71 -> 74.22 | Improved |
| add a fixed third camera view to Drive-π0 (ablation) | Drive-π0 · Bench2Drive (closed-loop) | Driving Score: 55.85 -> 62.38 (FL), 61.52 (FR), 63.26 (B) | Improved |
| add many fixed camera views simultaneously (ablation) | Drive-π0 · Bench2Drive (closed-loop) | Driving Score: 64.92 (3 views) -> 62.27 (6 views) | Mixed |
| use trajectory-level instead of token-level action router (ablation) | DriveMoE Action MoE · Bench2Drive (closed-loop) | Driving Score: 65.62 -> 73.88 | Improved |
| add skill supervision to the action router (ablation) | DriveMoE Action MoE · Bench2Drive (closed-loop) | Driving Score: 70.38 -> 74.22 | Improved |
| add Vision MoE (ablation) | DriveMoE · Bench2Drive (closed-loop) | Driving Score: 68.68 -> 74.22 | Improved |
| add Action MoE (ablation) | DriveMoE · Bench2Drive (closed-loop) | Driving Score: 67.31 -> 74.22 | Improved |
| add more shared experts (3 or 5 instead of 1) (ablation) | DriveMoE Action MoE · Bench2Drive (closed-loop) | Driving Score: 74.22 -> 73.88 / 73.46 | No clear effect |
| increase non-shared experts to 13 or 44 (ablation) | DriveMoE Action MoE · Bench2Drive (closed-loop) | Driving Score: 74.22 -> 70.88 / 68.22 | Hurt |
| add a sixth expert for ParkingExits (ablation) | DriveMoE Action MoE · Bench2Drive (closed-loop) | Driving Score: 73.81 -> 74.22 | Improved |
| DriveMoE vs Drive-π0 on open-loop L2 | DriveMoE · Bench2Drive (open-loop) | Avg. L2: 1.13 -> 0.96 (token), 1.01 (traj) | Mixed |
Closest papers by meaning
- DriveSpatial: A Benchmark for Spatiotemporal Intelligence in VLMs for Autonomous Driving (NeurIPS 2026 · similarity 0.90)
- AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning (NeurIPS 2025 · similarity 0.90)
- Less Language, More Latents: Annotation-Efficient VLAs for Driving (NeurIPS 2026 · similarity 0.89)
- DriveGPT4-V2: Harnessing Large Language Model Capabilities for Enhanced Closed-Loop Autonomous Driving (CVPR 2025 · similarity 0.89)
- SOLVE: Synergy of Language-Vision and End-to-End Networks for Autonomous Driving (CVPR 2025 · similarity 0.89)
- DriveCombo: Benchmarking Compositional Traffic Rule Reasoning in Autonomous Driving (CVPR 2026 · similarity 0.89)
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.
