# XMimic与XSoma实现人形机器人运动策略分钟级训练

> 原文链接：https://robotics.hellobike.com/papers/2026-11-17-2611-20452-43cdab15
> 频道：papers | 更新日期：2026-11-17 | 标签：auto, arxiv
> 核心摘要：基于JAX编译与多GPU并行优化SuperMimic框架，XMimic在4张H100上0.86分钟完成Unitree G1全动作跟踪，XSoma以1/8环境步数实现地形自适应，训练效率较基线提升3至12倍。

# XMimic与XSoma实现人形机器人运动策略分钟级训练

::: v-pre
> 本文由 AI 根据公开内容整理，可能有误，请以原文为准：[XMimic and XSoma: Learning Humanoid Motion Tracking and Locomotion in Under a Minute](https://arxiv.org/abs/2611.20452)（arXiv 2611.20452）

## 核心内容

针对人形机器人运动模仿与地形适应训练耗时痛点，研究提出XMimic和XSoma算法。两者继承SuperMimic世界模型解析梯度优化范式，通过JAX全栈编译及MuJoCo Warp仿真器集成，将串行小算子链重构为多GPU并行执行。实测在4张H100 GPU集群上，XMimic仅用0.86分钟即完成Unitree G1对LAFAN1数据集40个动作片段的跟踪学习；XSoma则以0.77分钟达到Holosoma基准测试目标，显著缩短Sim-to-Real迭代周期。

技术层面，XMimic利用最多4张GPU并行化训练迭代步骤，相比单卡BlinkMimic提速3.7倍。XSoma进一步解耦动力学建模，仅保留机器人本体动力学进入世界模型，足端地形高度等状态精确计算，采用1秒预测时域且摒弃Critic网络。该架构使XSoma相比FastSAC算法训练速度提升12.3倍，所需环境交互步数减少至原来的1/8，验证了编译型仿真与解析梯度结合在具身控制任务中的工程效能。

## 关键看点

- **训练加速**：XMimic在4×H100上0.86分钟完成G1全40个LAFAN1动作跟踪，比单卡BlinkMimic快3.7倍
- **样本效率**：XSoma达Holosoma基准仅需FastSAC八分之一环境步数，0.77分钟达标，提速12.3倍
- **系统架构**：基于JAX编译完整训练循环并集成MuJoCo Warp，将串行算子链重构为4卡并行执行
- **算法设计**：XSoma仅将机器人动力学纳入世界模型，足端地形精确计算，采用1秒时域无Critic优化

## 工程与产业观察

该工作证明了JAX编译型仿真与解析梯度结合可突破人形机器人RL训练算力瓶颈。通过算子并行化与世界模型解耦，将原本数小时的运动策略训练压缩至分钟级，大幅降低Sim-to-Real试错成本。这对依赖高频迭代的具身智能算法研发具有直接工程参考价值，提示开发者关注编译器级仿真加速与梯度计算优化。

## 英文摘要 (Abstract)

We present XMimic and XSoma, which learn humanoid motion tracking and locomotion on rough terrain in under a minute of training. Both follow SuperMimic, which learns a world model of the robot and optimizes the policy with analytic gradients through short rollouts imagined in this model. Existing implementations of this method do not make use of larger GPUs, as every training iteration is a long chain of small GPU operations that run one after another. XMimic and XSoma therefore shorten this chain and run the steps of every iteration at the same time on up to four GPUs. We compile the complete training loop with JAX, including the MuJoCo Warp simulator. For locomotion, XSoma leaves only the dynamics of the robot to the world model and computes everything else exactly, including the terrain height under the feet. It learns with a horizon of one second and without a critic. On four H100 GPUs, XMimic learns to track all 40 LAFAN1 clips with the Unitree G1 in 0.86 minutes, 3.7 times faster than BlinkMimic on one GPU. XSoma reaches the tracking target that Holosoma's tests set for FastSAC in 0.77 minutes, 12.3 times faster than FastSAC on one GPU, while needing only an eighth of its environment steps.

作者：Nico Bohlinger, Jan Peters
:::

## 原始参考来源 (Sources)

- [arXiv 2611.20452：XMimic and XSoma: Learning Humanoid Motion Tracking and Locomotion in Under a Minute](https://arxiv.org/abs/2611.20452)