Fast-dDrive: Efficient Block-Diffusion VLM for Autonomous Driving

arXiv:2605.23163 · 2026-05-28 공개 · arXiv · PDF

vla kv-cache autonomous-driving speculative-decoding throughput-optimization nuscenes block-diffusion trajectory-planning

Abstract

End-to-end autonomous driving via Vision-Language-Action (VLA) models demands a precarious balance between high-fidelity trajectory planning and efficient inference. Existing paradigms typically fall short: autoregressive (AR) VLAs are memory-bandwidth-bound on edge hardware and prone to exposure-bias drift, while full-sequence diffusion models preclude KV-cache reuse and suffer from "logical leakage" that violates the fundamental perceive-then-plan causality. We present Fast-dDrive, a block-diffusion VLA that performs bidirectional refinement within semantic units while enforcing strict causal ordering across them. Leveraging the observation that driving VLAs often emit structured JSON-like outputs, Fast-dDrive freezes structural tokens into a section scaffold and employs a section-aware training recipe that prioritizes safety-critical planning. We further introduce Scaffold Speculative Decoding to achieve AR-equivalent quality at significantly higher throughput. Finally, we propose a low-overhead test-time scaling scheme: by forking N stochastic trajectory rollouts from a single shared-prefix KV cache and averaging them, we effectively suppress prediction variance at a fractional computational cost. Empirical results demonstrate that Fast-dDrive redefines the speed-accuracy frontier for driving agents. On the WOD-E2E test set, Fast-dDrive achieves SOTA ADE@3s and ADE@5s, alongside the highest RFS among diffusion-based VLAs; on nuScenes, it reduces average L2 error to 0.32m (a 22% improvement). When integrated with SGLang, our framework delivers 12times throughput speedup over the AR baseline, narrowing the gap between high-capacity VLAs and the efficiency demands of real-time on-vehicle deployment.

한국어 요약

📋 한 줄 요약

**[자율주행 VLA / Block Diffusion]** Fast-dDrive가 AR·full-sequence diffusion의 한계를 block-diffusion으로 통합 — section 단위 양방향 정제·section 간 인과 정렬, scaffold speculative decoding·N rollout 평균화로 WOD-E2E ADE SOTA·nuScenes L2 0.32m·SGLang 결합 시 AR 대비 12× throughput.

🎯 핵심 기여도

💡 핵심 아이디어

자율주행 VLA의 속도·정확도 frontier는 AR과 diffusion을 block-diffusion으로 묶고, section 내부는 양방향·section 사이는 strict causal로 perceive-then-plan을 보존하면서 structured JSON scaffold·speculative decoding·shared-prefix N rollout 평균으로 효율과 안전을 동시 push하는 데서 새로 정의된다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: AR·diffusion 양쪽 paradigm의 핵심 한계를 block-diffusion으로 통합 해결, perceive-then-plan causality 보존이 안전성 보장, 12× throughput으로 on-vehicle 배포 실용성 도달, JSON scaffold 활용의 영리한 도메인 지식 통합. **한계**: Structured JSON 출력 가정에 일부 의존, 12× 가속이 SGLang 결합 한정, sim-to-real·다양 driving 환경 일반화 추가 검증.

🚀 실용적 활용