arXiv 2606.202462026-06-23cs.ROcs.AI

微调视觉-语言-动作模型所需的层数比你想象的少

Finetuning Vision-Language-Action Models Requires Fewer Layers Than You Think

Center for AI Research, VinUniversity（VinUniversity人工智能研究中心）；VinRobotics；University of Arkansas（阿肯色大学）；Technical University of Denmark（丹麦技术大学）；Hanoi University of Science and Technology（河内科技大学）；KAIST（韩国科学技术院）；Monash University（莫纳什大学）；Oldenburg University（奥尔登堡大学）；DFKI（德国人工智能研究中心）；University of Stuttgart（斯图加特大学）；IMPRS-IS（国际马克斯·普朗克智能系统研究学院）；Stanford University（斯坦福大学）；Technische Universität Darmstadt（达姆施塔特工业大学）

Gia-Binh Nguyen, Trong-Bao Ho, Thien-Loc Ha, Khoa Vo, Philip Lund Møller, Quang T. Nguyen, Long Dinh, Tung M. Luu, Tuan Dam, Vu Duong, Trung Le, Nghi D. Q. Bui,… 展开作者

Gia-Binh Nguyen, Trong-Bao Ho, Thien-Loc Ha, Khoa Vo, Philip Lund Møller, Quang T. Nguyen, Long Dinh, Tung M. Luu, Tuan Dam, Vu Duong, Trung Le, Nghi D. Q. Bui, Minh Vu, Tran Nguyen Le, An Thai Le, Ngan Le, Daniel Sonntag, James Zou, Jan Peters, Duy M. H. Nguyen, Ngo Anh Vien

阅读 PDF

arXiv 页面 ↗HTML 原文 ↗

AI总结：

本文发现VLA模型存在层间表示冗余，提出无需训练的压缩方法，通过去除冗余层将模型深度减少50%，实现40-50%训练加速和30%推理加速，性能不变。

摘要：

在大规模视频-机器人数据集上预训练的视觉-语言-动作（VLA）模型彻底改变了机器人操作，但其数十亿参数架构在下游微调和实时推理过程中带来了巨大的计算负担。在这项工作中，我们揭示了这些连续控制基础策略（例如pi_0、GR00T-N1.5）的一个高度非平凡的结构特性：尽管在多样化的物理轨迹上训练，它们表现出严重的逐层表示冗余。为了利用这一点，我们引入了一个完全无需训练的结构压缩流程，避免了现有方法需要加载全尺寸模型来学习优化的令牌缩减或动态层选择器的需求。相反，仅通过使用中心核对齐的单次前向传递来识别冗余层特征，我们移除孪生层以永久压缩模型深度高达50%，涵盖VLM主干和连续控制策略头。这种精简架构的下游微调带来了双重加速效益：训练时间减少40-50%，实时推理速度提升高达30%，同时匹配或超越全尺寸基模型性能。我们在三个模拟基准（LIBERO、RoboCasa、SimplerEnv）和10个跨4种不同机器人实体的多样化真实世界操作任务上全面验证了我们的方法。这些结果证明，先进的VLA所需的层数远少于先前假设，为可扩展的机器人学习提供了一种高度计算高效的范式。

英文摘要：

Vision-Language-Action (VLA) models pre-trained on massive video-robot datasets have revolutionized robotic manipulation, yet their multi-billion parameter architectures impose prohibitive computational burdens during downstream fine-tuning and real-time inference. In this work, we reveal a highly non-trivial architectural characteristic of these continuous control foundation policies (e.g., pi_0, GR00T-N1.5): despite being trained on diverse physical trajectories, they exhibit severe layer-wise representational redundancy. To exploit this, we introduce a structural compression pipeline that is entirely training-free, bypassing the need of existing methods to load full-scale models to learn optimized token reductions or dynamic layer selectors. Instead, using only a single forward pass via Centered Kernel Alignment to identify redundant layer features, we remove twin layers to permanently compress the model depth by up to 50% across both the VLM backbone and the continuous control policy head. Downstream fine-tuning of this streamlined architecture yields a dual acceleration benefit: a 40-50% reduction in training time and up to 30% faster real-time inference, while matching or exceeding full-scale base model performance. We comprehensively validate our method across three simulation benchmarks (LIBERO, RoboCasa, SimplerEnv) and 10 diverse real-world manipulation tasks across 4 unique robotic embodiments. These results prove that advanced VLAs require significantly fewer layers than previously assumed, offering a highly compute-efficient paradigm for scalable robot learning.

机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。