Research map / Vision-language and multimodal / Robot action and navigation
Scene-Adaptive VLA: Efficient Autonomous Driving via Dynamic Layer Routing
YUJIA YANG, Quan Yuan, Fan Jiawei, Yang Li, Tiange Fu, Xiaoyuan Fu, Guiyang Luo, Jinglin Li
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a scene-adaptive VLA for autonomous driving that dynamically routes inputs through layers to reduce onboard inference compute while preserving driving capability and explainability.
Problem
Large VLA models strain onboard compute during real-time autonomous driving inference, increasing energy use and limiting vehicle range, so computational overhead must be reduced without hurting driving performance or explainability.
Approach
Scene-adaptive dynamic layer routing: the model selects which layers to execute depending on the driving scene, so simple scenes use less computation. Details beyond the title and abstract are not available.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add scene-adaptive dynamic layer routing to SimLingo | SimLingo · Bench2Drive | LLM TFLOPs per decision: 0.59 -> 0.27 | Improved |
| add scene-adaptive dynamic layer routing to SimLingo | SimLingo · Bench2Drive | Driving Score: 85.07 -> 84.87 | No clear effect |
| add scene-adaptive dynamic layer routing to SimLingo | SimLingo · Bench2Drive | Comfortness: 33.67 -> 54.36 | Improved |
| add scene-adaptive dynamic layer routing to SimLingo | SimLingo · DriveLM-hard VQA and Commentary | SPICE: VQA 56.77 -> 54.91; Commentary 38.04 -> 35.37 | Hurt |
| add scene-adaptive layer routing to ORION | ORION · Bench2Drive | Driving Score: 77.74 -> 75.88 | Hurt |
| add scene-adaptive routing to ORION | ORION · Bench2Drive | LLM TFLOPs per decision: 9.07 -> 4.77 | Improved |
| set fixed 100% budget with scene-aware tokens (ablation) | SimLingo+SA · Bench2Drive | Driving Score: 85.07 -> 86.83 | Improved |
| reduce fixed active budget from 70% to 40% (ablation) | SimLingo+SA · Bench2Drive | Driving Score: 85.44 -> 78.20 | Hurt |
| add budget allocation module (vs random budget) (ablation) | SimLingo+SA · Bench2Drive ablation | Driving Score / avg budget: DS 83.81 -> 84.87; avg budget 60.03 -> 44.64 | Improved |
| add speed deviation token (ablation) | SimLingo+SA · Bench2Drive ablation | Avg. budget (%): 71.89 -> 44.64 | Improved |
| add path deviation token (ablation) | SimLingo+SA · Bench2Drive ablation | Avg. budget (%): 66.70 -> 44.64 | Improved |
| add scene deviation token (ablation) | SimLingo+SA · Bench2Drive ablation | Avg. budget (%): 54.37 -> 44.64 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 3 on most shared results.
- SimLingo — SimLingo: Vision-Only Closed-Loop Autonomous Driving with Language-Action Alignment (CVPR 2025) · better on 4 of 8 shared results
- ORION — ORION: A Holistic End-to-End Autonomous Driving Framework by Vision-Language Instructed Action Generation (ICCV 2025) · better on 4 of 4 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 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
- FIVE-VLA: Fast and EffectIVE Closed-Loop Autonomous Driving with Recurrent Action Memory (NeurIPS 2026 · similarity 0.91)
- Impromptu VLA: Open Weights and Open Data for Driving Vision-Language-Action Models (NeurIPS 2025 · 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)
- LaST-VLA: Thinking in Latent Spatio-Temporal Space for Vision-Language-Action in Autonomous Driving (NeurIPS 2026 · similarity 0.89)
- Less Language, More Latents: Annotation-Efficient VLAs for Driving (NeurIPS 2026 · similarity 0.89)
- VLP: Vision Language Planning for Autonomous Driving (CVPR 2024 · 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.
