Skip to content

README_cn.md

Latest commit

 

History

History
107 lines (72 loc) · 4.44 KB

README_cn.md

File metadata and controls

107 lines (72 loc) · 4.44 KB
Raw

beyondAMP: One-Step Integration of AMP into IsaacLab

概览 (Overview) 🌍

beyondAMP 提供了一个统一的管道，将对抗性运动先验 (Adversarial Motion Priors, AMP) 集成到任何 IsaacLab 机器人设置中，只需进行最少的修改，并能完全兼容自定义的机器人设计。

🚀 快速设置 (Fast Setup)

cd beyondAMP
bash scripts/setup_ext.sh
# Downloads assets, robot configs, and installs dependencies

可选的 VSCode 工作区设置：

python scripts/setup_vscode.py

📌 如何使用 (How to Use)

快速开始 (Quick Start)

  • 基础环境 (Basic environment): source/amp_tasks/amp_tasks/amp
  • PPO 配置 (PPO config for G1 robot): source/amp_tasks/amp_tasks/amp/robots/g1/rsl_rl_ppo_cfg.py

训练可以通过以下命令启动：

scripts/factoryIsaac/train.py --task AMPG1_Basic --headless
# scripts/factoryIsaac/train.py --task AMPG1_Soft --headless
# scripts/factoryIsaac/train.py --task AMPG1_Hard --headless

要评估或可视化一个已训练的检查点 (checkpoint)：

scripts/factoryIsaac/play.py --headless --target <path to your ckpt.pt> --video --num_envs 32

数据集准备 (Dataset Preparation) 💾

The dataset follows the same structure and conventions used in BeyondMimic(whole_body_tracking). All motion sequences should be stored as *.npz files and placed under data/datasets/, maintaining a consistent directory layout with the reference pipeline.

For motion retargeting and preprocessing, GMR is recommended for generating high-quality retargeted mocap data. TrackerLab may be used to perform forward kinematics checks and robot-specific adjustments, ensuring the motions remain physically plausible for your robot model.

With these tools, the dataset organization naturally aligns with the conventions established in BeyondMimic(whole_body_tracking), enabling seamless integration with the AMP training pipeline.

遵循 BeyondMimic 的数据集管道 (Following the dataset pipeline of BeyondMimic):

  • 动作文件 (Motion files): 将 *.npz 放入 data/datasets/
  • 推荐工具 (Recommended tools):
    • GMR 用于重定向动作 (retargeted motion)
    • TrackerLab 用于正向运动学 (FK) 验证和机器人特定的预处理

AMP 集成细节 (AMP Integration Details)

  • AMP observation group added via a new amp observation config
  • RSL-RL integration: source/rsl_rl/rsl_rl/env/isaaclab/amp_wrapper.py
  • Default transition builder: source/beyondAMP/beyondAMP/amp_obs.py

有关完整的教程和自定义，请参阅 (For full tutorial and customization, see) docs/tutorial.md。

<details> <summary><strong>附加说明 (Additional Notes)</strong></summary>

  • 完全模块化的 AMP 观察构建器 (Fully modular AMP observation builder)
  • 兼容 IsaacLab 4.5+ (Compatible with IsaacLab 4.5+)
  • 专为跨机器人形态的快速实验而设计 (Designed for rapid experimentation across robot morphologies)

</details>

🙏 致谢 (Acknowledgement)

参考仓库 (Referenced Repositories)

仓库 (Repository)                                                           用途 (Purpose)                              
robotlib                     机器人配置 (Robot configurations)                    
assetslib                   资产存储 (Asset storage)                        
TrackerLab                 数据组织与重定向工具 (Data organization & retargeting tools)
AMP_for_hardware           AMP 实现参考 (AMP implementation reference)          
BeyondMimic       数据集格式与跟踪比较 (Dataset format & tracking comparison)  

📘 引用 (Citation)

@software{zheng2025@beyondAMP,
  author = {Ziang Zheng},
  title = {beyondAMP: One step unify IsaacLab with AMP.},
  url = {https://github.com/Renforce-Dynamics/beyondAMP},
  year = {2025}
}