跳转至

文章背景与核心概要

随着具身AI代理的发展，视觉语言模型（VLM）正日益成为其核心推理引擎。由于机器人执行任务本质上是一个迭代过程——每一个动作都会根本性地改变物理环境与状态——代理必须具备持续感知、推理并验证其进展的能力。

为了满足这些多方面的需求，作者推出了 Capek 0.5，这是一个围绕以执行为中心的具身能力分类法构建的新型具身视觉语言模型。Capek 0.5 没有将模型针对孤立的数据集或静态任务进行训练，而是根据具身能力在整个主动执行过程中的具体功能角色对其进行了系统化的结构设计。

Capek 0.5: An Execution-Centric Vision-Language Model for Embodied Intelligence

Authors: Ying Chen, Weizhen Li, Zhe Hu, Zhenjiang Li, Rui Jiang, Zhifeng Gu, Lihuang Fang, Jiangping Liu, Lei Yi, and Jie Chen
Submitted: August 7, 2026
Primary Subject: Artificial Intelligence (cs.AI)
arXiv Identifier: arXiv:2608.06756

📌 Summary

Vision-language models (VLMs) are increasingly functioning as the core reasoning engines for embodied AI agents. Because robot execution is inherently iterative—where every action fundamentally alters the physical environment and state—agents must continually perceive, reason, and verify their progress.

To meet these multifaceted demands, the authors present Capek 0.5, a novel embodied vision-language model organized around an execution-centric capability taxonomy. Rather than training models against isolated datasets or static tasks, Capek 0.5 structures embodied capabilities according to their specific functional roles throughout active execution.

🧭 以执行为中心的能力分类法 (The Execution-Centric Capability Taxonomy)

Capek 0.5 将操作能力细分为四个主要的家族：

  1. 空间推理（Spatial Reasoning）： 理解并映射物理维度、距离和布局。
  2. 时间理解（Temporal Understanding）： 追踪事件序列、随时间推移的运动以及动作进展。
  3. 动作引导（Action Guidance）： 生成准确的步骤和物理干预措施以操纵环境。
  4. 状态验证（State Verification）： 持续确认目标或子目标是否已成功达成。

🧭 The Execution-Centric Capability Taxonomy

Capek 0.5 breaks down operational capabilities into four primary families:

  1. Spatial Reasoning: Understanding and mapping physical dimensions, distances, and layouts.
  2. Temporal Understanding: Tracking sequences of events, motion over time, and action progression.
  3. Action Guidance: Generating accurate steps and physical interventions to manipulate the environment.
  4. State Verification: Continuously confirming whether goals or sub-goals have been successfully achieved.

⚙️ 方法论与训练范式 (Methodology & Training Paradigm)

Capek 0.5 的训练流程通过一个两步过程将专家模型桥接到一个统一的系统中：

  • 专家获取（Specialist Acquisition）： 在共享基础模型的支持下，利用由可验证奖励支持的强化学习，独立培养每个能力家族。
  • 模型整合（Consolidation）： 通过权重空间合并（weight-space merging）将独立的专家合并为一个推理时模型，随后进行路由策略空间蒸馏（routed policy-space distillation），以在不损失性能的前提下保留专业的各项能力。

⚙️ Methodology & Training Paradigm

The training pipeline of Capek 0.5 bridges specialist models into a unified system through a two-step process:

  • Specialist Acquisition: Each capability family is independently cultivated using reinforcement learning backed by verifiable rewards derived from a shared backbone model.
  • Consolidation: The independent specialists are merged into a single inference-time model via weight-space merging, followed by routed policy-space distillation to retain specialized competencies without performance degradation.

📊 评估与结果 (Evaluation & Results)

Capek 0.5 在两个不同的规模（2B 和 35B-A3B）上进行了实例化，并在三个维度进行了严格测试：

  1. 基准测试套件（Benchmark Suites）： 在综合标准基准以及新推出的专为状态验证设计的基准 Capek-StateBench 上进行评估。
  2. 能力保留研究（Capability Retention Studies）： 进行对照分析，证明统一后的检查点成功保留了全部四个专门的能力流，且性能损失极小并经过了量化。
  3. 闭环仿真（Closed-Loop Simulation）： 实用验证展示了其成功迁移到交互式、模拟的具身环境中进行实时任务执行的能力。

📊 Evaluation & Results

Capek 0.5 is instantiated at two different scales (2B and 35B-A3B) and rigorously tested across three dimensions:

  1. Benchmark Suites: Evaluated against comprehensive standard benchmarks alongside Capek-StateBench, a newly introduced benchmark designed specifically for state verification.
  2. Capability Retention Studies: A controlled analysis proving that the unified checkpoint successfully retains all four specialized capability streams with minimal, quantified losses.
  3. Closed-Loop Simulation: Practical validation demonstrating successful transfer to interactive, simulated embodied environments for real-time task execution.