FLOWER

01 动机Motivation

大型视觉-语言-动作（VLA）模型在通用机器人策略学习中潜力巨大，但现有方法面临两大核心瓶颈：训练成本极高（往往需要数千 GPU 小时）以及推理效率低下（高显存占用、低频率输出），严重限制了其在真实场景与低资源环境中的落地。Large Vision-Language-Action (VLA) models hold great promise for generalist robot policy learning, but existing approaches face two core bottlenecks: extremely high training cost (often thousands of GPU hours) and poor inference efficiency (high GPU memory usage, low output frequency), which severely limits their deployment in real-world scenarios and low-resource environments.

"We present FLOWER, an efficient VLA flow policy capable of pretraining a 950M parameter model in only 200 H100 GPU hours, while achieving competitive performance across 190 tasks spanning simulation and real-world benchmarks."

现有方法的痛点主要体现在两个层面：The pain points of existing methods show up at two levels:

• 早期融合（Early Fusion）：将视觉 token 直接拼接进 LLM，导致序列长度激增，计算开销随之大幅上升。Early Fusion: visual tokens are concatenated directly into the LLM, causing sequence length to grow sharply and computational cost to rise with it.
• 晚期融合（Late Fusion）：虽然计算高效，但跨模态的语义对齐较弱，任务成功率受限。Late Fusion: computationally efficient, but cross-modal semantic alignment is weak, which limits task success rate.

FLOWER 通过截断预训练 VLM 的中间层特征（而非使用完整输出），在减少 20–35% 参数量的同时保留关键的跨模态语义信息，从而实现了训练效率与策略性能的双赢。By truncating the pretrained VLM and using its intermediate-layer features (rather than the full output), FLOWER preserves the key cross-modal semantic information while reducing parameters by 20–35%, achieving a win-win between training efficiency and policy performance.

02 方法Method

FLOWER 的整体框架由三个模块组成：视觉-语言编码器（VLM）提供多模态上下文表示；Flow Transformer 通过 cross-attention 接收 VLM 的中间层特征并执行 flow matching 动作解码；Global-AdaLN 以高效的参数共享方式为每层提供动作空间特定的调制信号。FLOWER's overall framework consists of three modules: the vision-language encoder (VLM) provides multimodal contextual representations; the Flow Transformer receives the VLM's intermediate-layer features through cross-attention and performs flow matching action decoding; Global-AdaLN supplies action-space-specific modulation signals to every layer via efficient parameter sharing.

中间模态融合（Intermediate-Modality Fusion）Intermediate-Modality Fusion

传统 VLA 模型通常使用 VLM 的最终层输出（晚期融合）或将图像 token 直接插入 LLM 输入（早期融合）。FLOWER 则采用"中间截断"策略：对于编码器-解码器结构（如 Florence-2），移除整个解码器（减少 50% 参数）；对于仅解码器结构（如 LLaMA），丢弃最后 30% 的层（减少 20–35% 参数）。截断后的中间层特征通过 cross-attention 传入 Flow Transformer。Conventional VLA models typically use the VLM's final-layer output (late fusion) or insert image tokens directly into the LLM input (early fusion). FLOWER instead adopts an "intermediate truncation" strategy: for encoder-decoder architectures (such as Florence-2), it removes the entire decoder (a 50% parameter reduction); for decoder-only architectures (such as LLaMA), it drops the last 30% of the layers (a 20–35% parameter reduction). The truncated intermediate-layer features are passed into the Flow Transformer through cross-attention.

"The intermediate representation prunes between 30% and 50% of pretrained VLM layers, yielding a 20–35% parameter reduction while preserving rich cross-modal semantics."

消融实验证明，中间融合相比早期融合提升 61 个百分点（LIBERO-Long：93.4% vs 33.4%），相比晚期融合提升 21 个百分点，优势显著。Ablations show that intermediate fusion outperforms early fusion by 61 percentage points (LIBERO-Long: 93.4% vs 33.4%) and late fusion by 21 percentage points, a substantial margin.

Global-AdaLN 条件化Global-AdaLN Conditioning

标准的 AdaLN-Zero 为 Diffusion Transformer 的每一层分配独立的调制参数，随层数增多参数量线性增长。Global-AdaLN 则将调制权重全局共享：Standard AdaLN-Zero assigns independent modulation parameters to every layer of the Diffusion Transformer, so the parameter count grows linearly with depth. Global-AdaLN instead shares the modulation weights globally:

• 一个共享的调制网络生成全局调制向量；a shared modulation network generates a global modulation vector;
• 每层附加一个轻量级 LoRA 适配器，在全局信号基础上引入层特异性偏移；each layer adds a lightweight LoRA adapter that introduces layer-specific offsets on top of the global signal;
• 相比 naive AdaLN-Zero，参数量减少 20% 以上，同时保留对不同动作空间统计特性的适应能力。compared with naive AdaLN-Zero, this reduces parameters by more than 20% while retaining the ability to adapt to the statistics of different action spaces.

Flow Matching 动作解码Flow Matching Action Decoding

FLOWER 的动作解码头基于 flow matching 框架（而非 DDPM 扩散过程），通过学习从噪声分布到动作分布的连续 ODE 流实现动作生成。Flow Transformer 接收带噪声的动作序列作为输入，以 VLM 中间层嵌入为条件，经去噪后输出机器人动作序列。该框架支持灵活的采样步数，在推理时可以更少步数获得高质量动作，从而实现 311.04 Hz 的高频推理（RTX 4090）。FLOWER's action decoding head is based on the flow matching framework (rather than a DDPM diffusion process), generating actions by learning a continuous ODE flow from the noise distribution to the action distribution. The Flow Transformer takes a noisy action sequence as input, conditions on the VLM's intermediate-layer embeddings, and outputs the robot action sequence after denoising. The framework supports a flexible number of sampling steps, so high-quality actions can be obtained with fewer steps at inference, enabling high-frequency inference at 311.04 Hz (RTX 4090).

03 实验Experiments

FLOWER 在 10 个基准上进行了系统评测，涵盖仿真（CALVIN、LIBERO、SIMPLER、Aloha）和真实世界（厨房操作）任务，共 190 项任务。对比基线包括 OpenVLA、π0、Octo、Seer、VPP 等代表性 VLA 模型。FLOWER is systematically evaluated on 10 benchmarks covering simulation (CALVIN, LIBERO, SIMPLER, Aloha) and the real world (kitchen manipulation), for a total of 190 tasks. Baselines include representative VLA models such as OpenVLA, π0, Octo, Seer, and VPP.

CALVIN ABC→D 基准CALVIN ABC→D Benchmark

LIBERO 多任务基准LIBERO Multi-Task Benchmark

真实世界厨房实验Real-World Kitchen Experiments

推理效率对比（RTX 4090）Inference Efficiency Comparison (RTX 4090)

FLOWER 推理速度比 π0 快 8%，比 OpenVLA 快 5,007%；显存仅为 π0 的 27.6%，OpenVLA 的 12.7%。FLOWER's inference is 8% faster than π0 and 5,007% faster than OpenVLA; its GPU memory is only 27.6% of π0's and 12.7% of OpenVLA's.

消融实验Ablation Studies

关键消融结果（CALVIN ABC 与 LIBERO-Long 双任务）：Key ablation results (on both CALVIN ABC and LIBERO-Long):

• 融合策略：早期融合 57.1% / 33.4%；晚期融合 71.2% / 61.8%；中间融合（本文）89.5% / 93.4%——中间融合比早期融合高出 61 百分点。Fusion strategy: early fusion 57.1% / 33.4%; late fusion 71.2% / 61.8%; intermediate fusion (ours) 89.5% / 93.4%—intermediate fusion exceeds early fusion by 61 percentage points.
• 层剪枝比例：保留完整 VLM（不剪枝）66.3%；剪枝 30%（最优）72.1%；剪枝 50% 66.4%。Layer-pruning ratio: full VLM (no pruning) 66.3%; pruning 30% (best) 72.1%; pruning 50% 66.4%.
• Global-AdaLN：相比 naive AdaLN-Zero 参数量减少 20% 以上，同时保持任务成功率不下降。Global-AdaLN: more than 20% fewer parameters than naive AdaLN-Zero, with no drop in task success rate.

04 局限性Limitations