# Kimodo-Policy：运动生成预训练驱动人形机器人操作成功率达72%

> 原文链接：https://robotics.hellobike.com/papers/2026-11-10-2611-11554-89c5ed0f
> 频道：papers | 更新日期：2026-11-10 | 标签：auto, arxiv
> 核心摘要：提出Kinematic Action Models范式，将全身运动生成作为预训练目标。Kimodo-Policy仅92M可训参数，在HumanoidArena多任务成功率达72.14%，真机8项任务均值60%，验证运动先验对人形操作的有效性。

# Kimodo-Policy：运动生成预训练驱动人形机器人操作成功率达72%

::: v-pre
> 本文由 AI 根据公开内容整理，可能有误，请以原文为准：[Kimodo-Policy: Learning Kinematic Action Models for Humanoid Loco-Manipulation](https://arxiv.org/abs/2611.11554)（arXiv 2611.11554）

## 核心内容

针对人形机器人移动操作中VLA模型难以捕捉运动结构的问题，研究提出Kinematic Action Models（KAMs）范式，将全身运动生成作为直接预训练目标。基于该范式的Kimodo-Policy在大规模loco-manipulation预训练后，于HumanoidArena基准测试中取得72.14%的最高综合成功率，并在8项真实世界任务中实现60%平均成功率，证明了以运动生成为基础的动作学习路径可行性。

系统采用冻结的Kimodo文本到运动骨干网络，结合轻量级ControlNet视觉条件模块，总参数量678M其中仅92M可训练。架构通过根节点-躯干-手部层级协调移动与操作，利用运动历史条件与重复动作块预测衔接合成与控制闭环。消融实验证实预训练运动骨干对性能提升具有关键贡献，且在SIMPLE基准上保持竞争力，为具身智能提供了区别于语义视觉先验的运动学基础方案。

## 关键看点

- **KAMs范式**：将全身运动生成作为预训练目标，通过任务条件、视觉观测与执行反馈接地时序与跨体协调先验
- **轻量化微调**：基于冻结Kimodo骨干与ControlNet视觉适配，总参数678M仅92M可训练，降低端到端训练成本
- **层级控制架构**：采用根-躯干-手分层协调移动操作，结合运动历史条件与动作块预测实现合成到闭环控制衔接
- **实测性能验证**：HumanoidArena多任务成功率72.14%居首，真机8项任务均值60%，消融证实运动骨干核心作用

## 工程与产业观察

该工作为人形机器人策略学习提供了「运动优先」的新工程坐标。相比纯VLA路线，KAMs通过解耦运动生成与任务决策，以极低可训参数实现高成功率，提示产业界在具身数据稀缺场景下，应重视高质量运动生成模型的预训练价值，将其作为操作策略的高效初始化底座而非仅依赖端到端多模态大模型。

## 英文摘要 (Abstract)

Humanoid loco-manipulation requires continuous coordination of high-dimensional whole-body motion to realize task objectives. While vision-language-action models and world action models transfer semantic and visual-dynamics priors to robot control, these priors do not necessarily capture the structure of robot motion. We introduce Kinematic Action Models (KAMs), a paradigm that treats whole-body motion as a direct pretraining target and uses a pretrained motion generator as the primary action-generation backbone. KAMs ground learned temporal and cross-body coordination priors through task conditioning, visual observations, and execution feedback. We present Kimodo-Policy as an initial instantiation, combining a frozen Kimodo text-to-motion backbone with lightweight ControlNet-based visual conditioning. A root-body-hand hierarchy coordinates locomotion and manipulation, while motion-history conditioning and repeated action-chunk prediction bridge motion synthesis and closed-loop control. Following large-scale loco-manipulation pretraining, the policy achieves the highest overall success among compared policies on multi-task HumanoidArena (72.14%), competitive performance on SIMPLE, and 60% mean success across eight real-world tasks, with 678M parameters (92M trainable). Ablations support the contribution of the pretrained motion backbone. Together, these results provide initial evidence for motion generation as a complementary foundation for humanoid action learning.

作者：Yunheng Wang, Jiaxi Zhang, Wei Lu, Yuetong Fang, Taowen Wang, Yixiao Feng, Shuning Zhang, Junzhe Xu, Yichi Wang, Zizhao Yuan, Kailun Huang, Lusong Li, Mingqian Zhou, Zecui Zeng, Renjing Xu
:::

## 原始参考来源 (Sources)

- [arXiv 2611.11554：Kimodo-Policy: Learning Kinematic Action Models for Humanoid Loco-Manipulation](https://arxiv.org/abs/2611.11554)