本文由 AI 分析生成
建立時間： 2026-06-15 來源： https://arxiv.org/abs/2509.19012
Summary
A September 2025 survey of pure VLA models (vision + language → action, end-to-end) covering 300+ papers. Proposes a five-paradigm taxonomy: autoregression-based, diffusion-based, reinforcement-based, hybrid, and specialized. Frames VLA as a paradigm shift from policy-based control to generalized robotics.
2025 年 9 月對純 VLA 模型的調查（視覺+語言→動作，端到端），涵蓋 300+ 篇論文。提出五範式分類：自回歸、擴散、強化學習、混合和專門化。將 VLA 定位為從策略控制到通用機器人學的典範轉移。
Prerequisites
- Foundation VLMs — VLAs extend vision-language models with action heads; understanding CLIP, BLIP, and transformer-based VLMs is prerequisite
- Behavior cloning vs. RL — the autoregression and diffusion paradigms are variants of imitation learning; the RL paradigm fine-tunes these with environmental reward; distinguishing the learning signals matters for understanding each paradigm’s data requirements
- Diffusion models — the diffusion paradigm (π0, Diffusion Policy) requires understanding score-matching and denoising; the key advantage is multimodal action distribution modeling
Core Idea
“Pure” VLA means end-to-end: the same model processes vision + language and directly outputs robot actions, with no intermediate hand-crafted representations. The five paradigms differ in how the action output is modeled. Autoregression treats actions as discrete tokens (enabling transfer from LLM pretraining but discretizing continuous control). Diffusion models the action distribution as a denoising process (preserves continuity, handles multimodality, but inference is slow). RL fine-tuning refines the policy with online interaction reward. Hybrid and specialized methods combine or restrict these.
Key Findings
- The field is converging on diffusion as the preferred action head for contact-rich tasks; autoregression dominates for high-level planning
- RL fine-tuning of pretrained VLAs is an emerging paradigm with strong early results but high compute cost
- Benchmarks and simulation platforms are fragmented; cross-paper comparison is unreliable due to inconsistent evaluation protocols
- Datasets and benchmarks: Open X-Embodiment, RLBench, LIBERO, MetaWorld, CALVIN — each covers different task profiles with different transfer properties
Limitations
- Author-stated: Survey scope is “pure” VLA — hybrid architectures with explicit planners or world models are excluded
- Unstated: 300+ papers reviewed in a fast-moving field means some 2025 advances (especially post-September) are missed
Reproducibility
- Code: Survey; references to individual paper codebases
- Datasets: Surveys existing benchmarks; no new dataset
- Compute: Survey paper
Insights
The five-paradigm taxonomy is useful as a map of design decisions: given a task (contact-rich manipulation vs. long-horizon instruction following) and constraints (inference latency, training data, compute), the taxonomy guides which paradigm to prototype first. The emergence of RL-based VLA fine-tuning is the most significant development not present in earlier VLA surveys — it closes the loop between VLA pretraining and environment-specific performance optimization.
Connections
- Clippings-an-anatomy-of-vision-language-action-models-from-modules-to-milestones-and-chall — earlier VLA anatomy survey
- Clippings-from-human-videos-to-robot-manipulation-scalable-vla-survey — scaling VLA with human video data
- human-robot-interaction — HRI context for VLA deployment