如何部署到自己的机器人

[kaolazhijiage]

作者您好，非常感谢您的开源工程，我的问题是，如何将训练好的文件成功部署到自己的机器人上？请问是否有这个方面的开源？例如您使用的真机是Xbot机器人，Xbot真机部署侧的代码是否可以提供呢？

[QianZhao-eng]

Deploying a trained RL policy to a real robot follows the same structural logic as sim2sim.py, but replaces the simulated physics engine with actual hardware communication. Here is a practical breakdown:

The sim2sim.py script in the repository is your best reference template. The key loop has three stages:
​

1. Read observations from the robot's sensors (joint positions, velocities, IMU, base orientation, commands)
2. Run inference on the loaded JIT-exported policy (.pt file)
3. Send actions (target joint positions) back to the robot via SDK/hardware interface

[JiaheHuo]

您好，我基于灵足时代的关节模组和维特智能的姿态传感器设计了一款自己的人形机器人并编写了基于cpp的sim2real部署代码。目前遇到一个问题为：机器人在issacGym和MuJoCo里面都能实现稳定行走，但是将一致的kpkd和scale部署到机器人上时需要将吊绳略微收紧才能实现稳定的行走。尝试过卸下双臂等减重措施但并不奏效，仍需要略微收紧吊绳。完全放下吊绳的情况下机器人可能不走直线、抖动都有发生。制作urdf时都进行了质量标定等工作，检查过obs维度、clip、推理频率问题。请问应该从哪些方向进行改进呢。呢？

[QianZhao-eng]

This is a classic sim-to-real gap issue — the root cause could be anything from a software bug to inaccurate robot dynamic parameters, or something else entirely. At a minimum, it's worth calibrating the dynamic parameters and carefully tuning the KP/KD gains to ensure the simulated and real joint trajectories are reasonably well-aligned before drawing any further conclusions. For reference, see Figures 3 and 4 in the Humanoid-Gym paper.