跳转至

文章背景与核心概要

当前的视觉-语言-动作(VLA)模型(如 OpenVLA、\(\pi_0\)、RT-2 和 RDT-1B)大多作为单一整体系统运行,直接输出原始的运动指令或简短的动作片段,而未能将复杂的行为组织为可重用的抽象。因此,这些模型在长周期(long-horizon)任务上的表现往往会下降,且缺乏可解释性。

为了克服这些局限性,作者推出了 REFACTOR-VLA,这是一个专为类型化运动程序的无监督库学习而设计的“清醒/睡眠”(wake/sleep)系统。该系统通过清醒阶段生成类型化 lambda 项,并结合睡眠阶段利用行为等效核进行运动程序聚类,同时通过最小描述长度(MDL)和回报保持门控机制来精炼技能库。实验证明,该方法显著超越了现有的先进基线模型,为解决机器人具身智能的可解释性与长序贯规划难题提供了新途径。


REFACTOR-VLA: Unsupervised Library Learning of Typed Motor Programs

Authors: Riyaaz Shaik, Chandru Venkataraman
Published: September 1, 2026
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Robotics (cs.RO)
arXiv: 2609.01215 [cs.LG]

Authors: Riyaaz Shaik, Chandru Venkataraman
Published: September 1, 2026
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Robotics (cs.RO)
arXiv: 2609.01215 [cs.LG]


Executive Summary

Most current Vision-Language-Action (VLA) models (such as OpenVLA, \(\pi_0\), RT-2, and RDT-1B) operate as monolithic systems that emit raw motor commands or short action chunks without organizing behavior into reusable abstractions. Consequently, these models often degrade on long-horizon tasks and lack interpretability.

To overcome these limitations, the authors introduce REFACTOR-VLA, a wake/sleep system designed for unsupervised library learning of typed motor programs.

Key Innovations & Findings

  • Wake/Sleep Architecture:
  • Sleep Phase: Clusters motor-program fragments using a Behavioral-Equivalence Kernel (BEK) computed from rollouts of a learned latent world model (\(M_\phi\)).
  • Wake Phase: Emits typed lambda terms over a Hindley-Milner-inspired vocabulary, which are then consumed by a library-conditioned rectified-flow action decoder.
  • Gating Mechanisms: Abstractions are only admitted if they successfully pass Minimum Description Length (MDL) and return-preservation gates.
  • Model Capacity vs. Objective Quality: Scaling the world model from 188M to 430M parameters actually degraded performance across all 4 LIBERO suites, proving that capacity alone is insufficient. Instead, the training objective is paramount: incorporating an auxiliary supervised contrastive (InfoNCE) loss during world-model warmup drastically improves sleep-phase clustering.
  • Performance: REFACTOR-VLA outperforms the strongest published baseline across all 4 LIBERO suites by a mean \(\Delta\) of \(+0.184\), achieving Normalized Mutual Information (NMI) scores of:
  • Object: \(0.462 \pm 0.021\)
  • Spatial: \(0.867 \pm 0.025\)
  • Goal: \(0.915 \pm 0.013\)
  • LIBERO-10: \(0.754 \pm 0.010\)

执行摘要

大多数当前的视觉-语言-动作(VLA)模型(如 OpenVLA、\(\pi_0\)、RT-2 和 RDT-1B)作为单一整体系统运行,输出原始运动指令或短动作片段,而未能将行为组织为可重用的抽象。因此,这些模型在长跨度任务上经常性能退化,并且缺乏可解释性。

为了克服这些局限性,作者推出了 REFACTOR-VLA,这是一个专为类型化运动程序的无监督库学习而设计的清醒/睡眠(wake/sleep)系统。

核心创新与发现

  • 清醒/睡眠架构:
  • 睡眠阶段: 使用从学习到的潜在世界模型(\(M_\phi\))的轨迹展开(rollouts)计算出的行为等效核(BEK)对运动程序片段进行聚类。
  • 清醒阶段: 在受 Hindley-Milner 启发词汇表的基础上发射类型化 lambda 项,随后由库条件化的整流流(rectified-flow)动作解码器对其进行消费。
  • 门控机制: 只有成功通过最小描述长度(MDL)和回报保持门控的抽象才会被收录。
  • 模型容量与目标质量: 将世界模型从 1.88 亿参数扩展到 4.3 亿参数实际上导致了所有 4 个 LIBERO 套件性能的下降,这证明仅靠模型容量是不够的。相反,训练目标至关重要:在世界模型预热期间引入辅助监督对比(InfoNCE)损失,会大大改善睡眠阶段的聚类效果。
  • 性能表现: REFACTOR-VLA 在所有 4 个 LIBERO 套件上均超越了最强的已发布基线,平均提升 \(\Delta\)\(+0.184\),并取得了以下归一化互信息(NMI)得分:
  • 物体(Object): \(0.462 \pm 0.021\)
  • 空间(Spatial): \(0.867 \pm 0.025\)
  • 目标(Goal): \(0.915 \pm 0.013\)
  • LIBERO-10: \(0.754 \pm 0.010\)

元数据与参考链接


License: license icon View License

许可证: license icon 查看许可证