# VLA灵巧手微调：冻结VLM全训动作专家优于LoRA与SFT指标

> 原文链接：https://robotics.hellobike.com/papers/2026-11-19-2611-23676-e14cc272
> 频道：papers | 更新日期：2026-11-19 | 标签：auto, arxiv
> 核心摘要：研究提出适配LEAP Hand的VLA微调范式，证实SFT成功率与DSRL效果弱相关。冻结VLM并全量训练动作专家策略可引导性最强，LoRA表现最差，建议摒弃SFT Rollout作为检查点选择依据。

# VLA灵巧手微调：冻结VLM全训动作专家优于LoRA与SFT指标

::: v-pre
> 本文由 AI 根据公开内容整理，可能有误，请以原文为准：[Finetuning VLAs for Dexterous Manipulation: The Right Checkpoint Is All You Need](https://arxiv.org/abs/2611.23676)（arXiv 2611.23676）

## 核心内容

针对VLA模型向多指灵巧手迁移难及遥操作数据稀缺问题，研究基于 $\pi_{0.5}$ 模型与 LEAP Hand 硬件，在仿真环境中评估了五种微调方案。实验表明，监督微调（SFT）的成功率与后续扩散引导强化学习（DSRL）的性能相关性极低，最优 SFT 检查点往往并非最佳引导起点，揭示了传统指标在灵巧操作适配中的失效。

技术验证显示，冻结视觉语言模型（VLM）参数、仅对动作专家进行全量训练的策略，在训练集及分布外放置任务中均展现出最高的可引导性；相比之下，低秩自适应（LoRA）方法的可引导性最差。研究确立了“冻结 VLM + 全训动作专家”的适配范式，并明确建议不再使用 SFT Rollout 成功率作为模型检查点的筛选标准。

## 关键看点

- **指标解耦**：SFT阶段成功率与DSRL后训练效果呈弱相关，最优SFT检查点非最佳引导基座
- **最优架构**：冻结VLM参数并全量微调动作专家，在OOD场景下策略可引导性显著优于其他方案
- **LoRA局限**：低秩自适应方法在灵巧手VLA微调中可引导性最差，不适合作为核心适配手段
- **数据效率**：仅依赖少量演示数据即可完成从平行夹爪到LEAP Hand多指灵巧操作的策略迁移

## 工程与产业观察

该研究修正了具身大模型微调的工程直觉，证明在灵巧操作场景下，盲目追求SFT高成功率或采用轻量级LoRA反而限制上限。冻结VLM保全语义理解、全训动作专家以适配高维运动学，是解决多指手数据稀缺与Sim-to-Real鸿沟的高性价比路径，为VLA落地提供了关键调参基准。

## 英文摘要 (Abstract)

Vision-Language-Action models have demonstrated strong manipulation capabilities. A limitation of VLA models is that most manipulation datasets focus on parallel-jaw gripper systems, making transfer to multi-finger hands challenging, particularly as teleoperation datasets are harder to collect with multi-fingered robots. In this work, we study how to best adapt VLAs to dexterous hands by pairing supervised fine-tuning (SFT) with diffusion-steering reinforcement learning (DSRL). Using $\pi_{0.5}$ and a LEAP hand, we evaluate five fine-tuning recipes across two simulation tasks using only a handful of demonstrations. Our empirical analysis yields three key insights: (1) SFT success is poorly correlated with post-DSRL success, meaning the best SFT checkpoint is rarely the most effective to steer; (2) freezing the vision-language model (VLM) and tuning only the action expert produces the most steerable policy for both training and out-of-distribution placements; and (3) low-rank adaptation (LoRA) is the least steerable approach. Consequently, we recommend an adaptation recipe that freezes the VLM, fully trains the action expert, and eschews SFT rollout success as a checkpoint selection metric.

作者：Artur Kuramshin, Charlotte Cloutier, Michael Przystupa, Kirsty Ellis, Mohammad S. Akhter, Glen Berseth
:::

## 原始参考来源 (Sources)

- [arXiv 2611.23676：Finetuning VLAs for Dexterous Manipulation: The Right Checkpoint Is All You Need](https://arxiv.org/abs/2611.23676)