# CAPABLE框架实现冻结VLA策略关节故障在线自适应恢复

> 原文链接：https://robotics.hellobike.com/papers/2026-10-08-2610-11971
> 频道：papers | 更新日期：2026-10-08 | 标签：auto, arxiv
> 核心摘要：针对VLA模型关节故障失效问题，提出CAPABLE框架。通过自监督能力推断与残差强化学习，无需故障标签即可在线修正动作。在LIBERO任务中成功率从24.8%升至59.3%，并在Franka Panda真机验证有效。

# CAPABLE框架实现冻结VLA策略关节故障在线自适应恢复

::: v-pre
> 本文由 AI 根据公开内容整理，可能有误，请以原文为准：[CAPABLE: Capability-Aware Policy Adaptation via Behavioral Latent Encoding](https://arxiv.org/abs/2610.11971)（arXiv 2610.11971）

## 核心内容

针对视觉-语言-动作（VLA）模型在关节故障下物理执行失效且缺乏通用恢复手段的痛点，研究提出CAPABLE统一适配框架。该方法专为冻结参数的VLA设计，集成自监督能力推断与残差强化学习，无需任务重训、故障标签或特权本体信息。实验显示，在28个LIBERO任务中，针对训练集未包含的执行器故障，CAPABLE将任务成功率从基线的24.8%显著提升至59.3%，并在六关节留一法测试及`Franka Panda`真机上验证了跨故障类型的泛化恢复能力。

技术层面，CAPABLE利用共享时序编码器、雅可比矩阵接地（Jacobian grounding）、跨关节注意力机制及自监督物理预测，从指令-响应历史与运动学中在线推断各关节实际运动贡献度。生成的“能力表征”作为条件输入残差策略，为VLA机械臂动作添加有界修正量。相比参数匹配的全局历史基线，该方案性能高出17.4个百分点，且在健康状态下保持原有性能，实现了无需显式故障诊断的隐式自适应控制，解决了具身智能体在非结构化环境下的鲁棒性难题。

## 关键看点

- **在线能力推断**：基于命令响应历史与运动学，利用雅可比矩阵接地与跨关节注意力机制，实时量化各关节对末端执行器的实际运动贡献度
- **残差策略修正**：以推断的能力表征为条件，通过残差强化学习为冻结VLA输出有界动作修正量，无需故障标签或特定关节标识符
- **仿真性能跃升**：在28个LIBERO任务中，针对未见执行器故障的成功率从24.8%提升至59.3%，优于全局历史基线17.4个百分点
- **真机泛化验证**：通过六关节留一法实验证实跨执行器迁移能力，并在物理Franka Panda机器人上成功实现未知故障家族的自适应恢复

## 工程与产业观察

CAPABLE为具身大模型落地提供了关键的容错工程范式。其核心价值在于解耦了“策略训练”与“硬件维护”，使冻结参数的VLA能在关节退化或损坏时通过残差策略在线补偿，避免了昂贵的全量重训。这种基于运动学接地与自监督推断的隐式适配机制，显著提升了机器人在非结构化环境中的长周期运行鲁棒性，对降低具身智能运维成本具有直接参考价值。

## 英文摘要 (Abstract)

Vision-language-action (VLA) policies assume the embodiment on which they were trained and can fail when a joint fault changes how commanded actions are physically executed. Existing fault-recovery methods often require task-specific retraining, fault labels, explicit diagnosis, or privileged embodiment information. We introduce CAPABLE, a unified capability-aware adaptation framework for frozen VLAs that integrates self-supervised capability inference with residual reinforcement learning. CAPABLE infers capability, how much of the commanded motion each joint actually realizes and how that motion contributes to end-effector behavior, online from command-response history and kinematics using a temporal encoder shared across joints, Jacobian grounding, cross-joint attention, and self-supervised physical prediction. The resulting representation conditions a residual policy that adds bounded corrections to the VLA arm action without fault labels or faulty-joint identifiers. Across 28 LIBERO tasks, CAPABLE raises success on an actuator excluded from fault training from 24.8% to 59.3%, outperforming a parameter-matched global-history baseline by 17.4 points while preserving healthy performance. Leave-one-actuator-out experiments across six joints show that this transfer is not specific to one actuator, and additional evaluations characterize transfer to unseen fault families and demonstrate recovery on a physical Franka Panda. https://capable-vla.github.io/

作者：Mohammad Khoshnazar, Mohammad Dehghani Tezerjani, Deyuan Qu, Zhiyuan Gao, Yanxiang Zhan, Jeroen Schafer, Andrew Melnik, Qing Yang, Michael Beetz
:::

## 原始参考来源 (Sources)

- [arXiv 2610.11971：CAPABLE: Capability-Aware Policy Adaptation via Behavioral Latent Encoding](https://arxiv.org/abs/2610.11971)