VLA-Adapter: An Effective Paradigm for Tiny-Scale Vision-Language-Action Model

01 动机Motivation

现有 VLA 模型普遍依赖数十亿参数的视觉语言骨干（如 7B LLM）与大规模机器人预训练数据，导致训练成本高、推理延迟大，难以在消费级硬件上部署。本文的核心问题是：能否仅用一个微小的 VLM 骨干（0.5B），通过设计更好的 bridging 方式，达到甚至超越大模型的操作性能？Existing VLA models generally rely on vision-language backbones with billions of parameters (e.g. a 7B LLM) together with large-scale robotic pre-training data, which makes training expensive and inference slow, and hinders deployment on consumer-grade hardware. The central question of this paper is: can a tiny VLM backbone (0.5B) alone, through a better-designed bridging scheme, match or even surpass the manipulation performance of large models?

"We propose VLA-Adapter to reduce VLA's reliance on large-scale Vision-Language Models and extensive pre-training … achieving state-of-the-art performance using only a 0.5B-parameter backbone without robotic data pre-training."

VLA-Adapter teaser：参数量 vs 性能对比 — 图1：VLA-Adapter 的核心特点——使用最小规模骨干（0.5B），在 LIBERO 系列基准上达到 SOTA 性能，同时具备极高推理吞吐量。横轴为骨干参数量，纵轴为平均成功率。Figure 1: The defining characteristic of VLA-Adapter — with the smallest-scale backbone (0.5B) it reaches SOTA performance on the LIBERO benchmark suite while retaining extremely high inference throughput. The horizontal axis is backbone parameter count, the vertical axis average success rate.

0.5B骨干参数量（vs 同类 7B）Backbone parameters (vs. 7B peers)

97.3%LIBERO 平均成功率LIBERO average success rate

219.2 Hz推理吞吐量（OpenVLA 仅 4.2 Hz）Inference throughput (OpenVLA: only 4.2 Hz)

8 h单块消费级 GPU 完整训练时长Full training time on a single consumer-grade GPU

02 方法Method

VLA-Adapter 的核心是将 VLM 产生的多模态表征以最优方式"桥接"到动作空间。论文首先系统梳理了四种 bridging paradigm（图2），随后提出包含条件探索与 Bridge Attention 设计的 Policy 模块（图3），并通过 97M 参数的轻量网络（图4）输出 H 步 action chunk。The core of VLA-Adapter is to “bridge” the multimodal representations produced by the VLM into the action space in an optimal way. The paper first systematically reviews four bridging paradigms (Figure 2), then proposes a Policy module that combines condition exploration with a Bridge Attention design (Figure 3), and finally emits an H-step action chunk through a lightweight network of 97M parameters (Figure 4).

现有 bridging paradigm 对比 — 图2：四种已有 Vision-Language→Action bridging 范式的系统梳理。从左至右分别对应 token concatenation、cross-attention injection、prefix conditioning 和 full fine-tuning 等典型策略，揭示各自的信息传递方式与局限。Figure 2: A systematic review of four existing Vision-Language→Action bridging paradigms. From left to right they correspond to typical strategies such as token concatenation, cross-attention injection, prefix conditioning and full fine-tuning, revealing the information pathway and the limitations of each.

图3：VLA-Adapter 整体框架。冻结的 VLM 骨干提取 raw latent features ℛ 与 ActionQuery latent AQ；Policy 模块通过 Bridge Attention 将这两类条件注入动作空间，输出 H 步动作序列。Figure 3: The overall VLA-Adapter framework. A frozen VLM backbone extracts the raw latent features ℛ and the ActionQuery latent AQ; the Policy module injects these two kinds of conditions into the action space through Bridge Attention and outputs an H-step action sequence.

条件分析：哪层特征最有用？Condition analysis: which layer of features is most useful?

论文对 VLM 内部特征进行系统分析，得出三个关键结论：The paper systematically analyses the internal features of the VLM and draws three key conclusions:

Raw features：中间层特征优于深层特征——深层特征更偏向语义抽象，缺乏对动作执行至关重要的细粒度多模态细节。Raw features: intermediate-layer features outperform deep-layer ones — deep features lean towards semantic abstraction and lack the fine-grained multimodal detail that is essential for action execution.
ActionQuery：深层 ActionQuery 优于浅层——ActionQuery 是任务导向的、经过训练的聚合表征，越深层越富有多模态信息。ActionQuery: deep-layer ActionQuery outperforms shallow-layer ActionQuery — ActionQuery is a task-oriented, trained aggregate representation, and the deeper the layer, the richer its multimodal information.
多层组合：多层特征的融合始终优于单层特征，最终采用"all-layer raw + deep ActionQuery"的组合。Multi-layer combination: fusing features from multiple layers is consistently better than any single layer; the final design adopts the “all-layer raw + deep ActionQuery” combination.

Bridge Attention 设计The Bridge Attention design

Bridge Attention 由两个 cross-attention 和一个 self-attention 构成。它分别处理 raw latent ℛ（多模态细节）与 ActionQuery latent AQ（聚合信息），通过 MLP 生成 key-value pair。一个可学习参数"Ratio g"以 tanh(g) ∈ [−1, 1] 控制 raw features 的注入程度，防止分布不稳定；AQ 则以固定权重 1.0 注入，从而在灵活性与稳定性间取得最优平衡。Bridge Attention consists of two cross-attention modules and one self-attention module. It processes the raw latent ℛ (multimodal detail) and the ActionQuery latent AQ (aggregated information) separately, producing key-value pairs through an MLP. A learnable parameter “Ratio g” uses tanh(g) ∈ [-1, 1] to control how strongly raw features are injected, preventing distributional instability, while AQ is injected with a fixed weight of 1.0 — striking the best balance between flexibility and stability.

Policy with Bridge Attention 架构 — 图4：Policy 模块详细结构（97M 参数）。每一层包含 Bridge Attention 和 Feed-Forward Network，以 {ℛ, AQ, 初始动作 latent, 本体感知状态} 为输入，输出 H 步 action chunk。Figure 4: Detailed structure of the Policy module (97M parameters). Every layer contains a Bridge Attention and a Feed-Forward Network, taking {ℛ, AQ, initial action latent, proprioceptive state} as input and outputting an H-step action chunk.

03 实验Experiments

实验在 LIBERO（Spatial / Object / Goal / Long 四个子集）和 CALVIN ABC→D 两大基准上与多个 SOTA VLA 对比，同时评估效率指标（吞吐量、延迟）和冻结骨干场景下的泛化能力。骨干使用 Qwen2.5-0.5B，无机器人数据预训练。Experiments compare VLA-Adapter with several SOTA VLAs on two major benchmarks — LIBERO (the Spatial / Object / Goal / Long subsets) and CALVIN ABC→D — and also evaluate efficiency metrics (throughput, latency) and generalisation under a frozen backbone. The backbone is Qwen2.5-0.5B, with no robotic-data pre-training.

LIBERO 基准对比（表5）LIBERO benchmark comparison (Table 5)

方法Method	参数量Params	Spatial	Object	Goal	Long	Avg.
π0	3B	96.8%	98.8%	95.8%	85.2%	94.2%
SmolVLA	2.2B	93.0%	94.0%	91.0%	77.0%	88.8%
OpenVLA-OFT	7B	97.6%	98.4%	97.9%	94.5%	97.1%
VLA-Adapter（ours）VLA-Adapter (ours)	0.5B	97.8%	99.2%	97.2%	95.0%	97.3%
VLA-Adapter-Pro（ours）VLA-Adapter-Pro (ours)	0.5B	99.6%	99.6%	98.2%	96.4%	98.5%

CALVIN ABC→D 基准对比（表6）CALVIN ABC→D benchmark comparison (Table 6)

方法Method	参数量Params	Task 1	Task 2	Task 3	Task 4	Task 5	Avg. Len
OpenVLA-OFT	7B	96.3%	89.1%	82.4%	75.8%	66.5%	4.10
VLA-Adapter（ours）VLA-Adapter (ours)	0.5B	99.1%	94.6%	88.8%	82.8%	76.5%	4.42
VLA-Adapter-Pro（ours）VLA-Adapter-Pro (ours)	0.5B	98.5%	95.0%	90.5%	85.3%	80.0%	4.50

推理效率对比（表4）Inference efficiency comparison (Table 4)

方法Method	吞吐量 (Hz)Throughput (Hz)	延迟 (sec)Latency (sec)
OpenVLA	4.2	0.2396
VLA-Adapter（ours）VLA-Adapter (ours)	219.2	0.0365

VLA-Adapter 的吞吐量比 OpenVLA 提升 52×，延迟降低 6.5×，满足实时控制需求。VLA-Adapter raises throughput over OpenVLA by 52× and cuts latency by 6.5×, meeting the requirements of real-time control.

真实机器人实验Real-robot experiments

真实机器人任务对比 — 图5：在 Synria Alicia-D 机器人系统上进行真实世界任务评估，涵盖多种场景（抓取、摆放、工具使用等），VLA-Adapter 表现出稳健的迁移能力。Figure 5: Real-world task evaluation on the Synria Alicia-D robot system, covering a range of scenarios (grasping, placing, tool use and so on); VLA-Adapter shows robust transfer capability.

消融实验Ablation studies

消融实验（表7，LIBERO-Long）验证了各条件组合的贡献：The ablation study (Table 7, LIBERO-Long) verifies the contribution of each condition combination:

条件组合Condition combination	成功率Success rate
仅 last-layer rawlast-layer raw only	85.8%
仅 ActionQueryActionQuery only	90.2%
仅 intermediate rawintermediate raw only	88.4%
仅 all-layer rawall-layer raw only	90.6%
all-layer raw + ActionQuery（ours）all-layer raw + ActionQuery (ours)	95.0%

注入度消融（表8）显示：raw features 采用可学习 tanh(g)、AQ 固定为 1.0 的组合最优（95.0%），优于两者均可学习（92.6%）或均固定（91.4%）的方案。ActionQuery token 数量实验（图8）表明 64 个 token 为最优平衡点，更多 token 带来冗余干扰。The injection-strength ablation (Table 8) shows that a learnable tanh(g) for raw features combined with AQ fixed at 1.0 is optimal (95.0%), beating both being learnable (92.6%) and both being fixed (91.4%). The experiment on the number of ActionQuery tokens (Figure 8) indicates that 64 tokens is the best trade-off point, since more tokens bring redundant interference.

骨干缩放实验（表2）进一步表明，VLA-Adapter 框架对骨干规模并不敏感：从 0.5B 到 7B，LIBERO-Long 成功率仅从 95.0% 提升至 95.4%，验证了框架本身的有效性而非单纯依赖骨干能力。The backbone-scaling experiment (Table 2) further shows that the VLA-Adapter framework is insensitive to backbone scale: going from 0.5B to 7B lifts the LIBERO-Long success rate only from 95.0% to 95.4%, which confirms the effectiveness of the framework itself rather than a mere reliance on backbone capacity.

04 局限性Limitations

Note: 以下局限性均为论文作者在 Limitations 章节中明确陈述（stated）。All limitations below are explicitly stated by the authors in the Limitations section of the paper.

真实场景泛化能力有待提升Generalisation to real-world scenarios needs improvement

"Because VLA-Adapter is not pre-trained on a large amount of embodied data and the scale is tiny, its generalization in real-world systems needs to be improved."——受限于骨干规模极小（0.5B）且未在大规模具身数据上预训练，模型在新场景、新任务上的零样本或少样本泛化能力仍有差距。“Because VLA-Adapter is not pre-trained on a large amount of embodied data and the scale is tiny, its generalization in real-world systems needs to be improved.” — constrained by an extremely small backbone (0.5B) and the absence of pre-training on large-scale embodied data, the model still falls short in zero-shot and few-shot generalisation to new scenes and new tasks.

动作质量依赖 VLM 条件质量Action quality depends on the quality of the VLM conditions

"The quality of the actions generated depends on conditions provided by the VLM and how they are used."——Bridge Attention 的效果上限受 VLM 提供的 raw features 和 ActionQuery 特征质量约束，VLM 骨干的表达能力直接影响 Policy 能否获得足够有效的条件信号。“The quality of the actions generated depends on conditions provided by the VLM and how they are used.” — the performance ceiling of Bridge Attention is bounded by the quality of the raw features and ActionQuery features supplied by the VLM; the expressive power of the VLM backbone directly determines whether the Policy can obtain sufficiently effective conditioning signals.

训练流程仍较简单，未探索强化学习The training pipeline is still simple; reinforcement learning is unexplored

"The fundamental training process is still relatively simple, and processes such as reinforcement learning can be explored."——当前仅使用模仿学习（行为克隆）范式训练，未引入 RL fine-tuning 或 RLHF 等更复杂的训练策略，存在进一步提升空间。“The fundamental training process is still relatively simple, and processes such as reinforcement learning can be explored.” — the model is currently trained under a pure imitation-learning (behaviour cloning) paradigm, without more elaborate strategies such as RL fine-tuning or RLHF, so there is room for further improvement.