During `play_amp_animation`, the expert's animation jitters. Does this affect the expert's data performance and thus the training?

[JinhaoGao]

When using the Tienkung project, we found that expert data has a significant impact on training results. This includes, but is not limited to, performing ablation experiments on expert data alone, which revealed that some training methods failed, while others trained quickly and effectively. However, we are currently unsure how to determine the quality of the expert data. Visualizing the motion, we noticed that the expert motions provided by Tienkung are jittery. Is this normal?

tienkung.run.expert.motion-2026-04-28_20.01.21.mp4

tienkung.walk.expert.motion-2026-04-28_20.00.09.mp4

[BryceWang0527]

2026-05-11.15-15-51.mp4

[BryceWang0527]

This is not expected behavior. We checked the expert motions and did not observe noticeable jittering on our side. Could you please pull the latest updates from the repository and try again?

[JinhaoGao]

或者这是否与5.0.0版本的isaaclab有关系呢？并且还有个问题想请教您，我们发现策略输出的action中维度是我们机器人的关节维度，即20dof，但是在play.py中print出来的顺序（joints name)却不是我们期待的顺序，并且没有找到输出action每个纬度对应关节值是在哪里定义的，貌似由isaaclab内部自行定义么，与我们的urdf中关节顺序和assets/robot.py定义的顺序均不一致，只能够print出来才能确定action的每个维度对应的关节么，以便在sim2sim过程中完成关节映射。期待您的回答，十分感谢。（附件为dev分支，dex walk任务的play结果，与tienkung-dex机器人的urdf和assets中的py文件顺序也不相同，即hip roll/pitch/yaw ....，这令我们很费解）

[JinhaoGao]

This is not expected behavior. We checked the expert motions and did not observe noticeable jittering on our side. Could you please pull the latest updates from the repository and try again?

谢谢您的回答，isaaclab索引问题我找到了答案，是广度优先搜索。不过专家数据部分我们确实没有找到答案，虽然目前播放出来的动作都是卡顿有重影的，但是不同的动作确实会很大程度影响训练效果，如训练2w轮却只会站立，到换了专家动作，却很轻松学会行走，我做过对照组实验，确定是专家数据的问题，但是由于播放时候一致性的抖动，让我们无法判断是什么原因导致的问题，期待您的回答。（我们是自研机器人，通过GMR重定向的AMASS SFU 0007 walking动作，仅仅是团队不同人重定向出的同一个动作，并非是不同的原始动作，这令我们更加困惑）

[BryceWang0527]

Thanks for your update. I get there are actually two separate issues here.
First, the motion jitter in playback might be fixed by aligning your IsaacLab/IsaacSim version with the official repo. You can also try only rendering without stepping the simulation to locate the problem.
Second, your GMR retargeting pipeline itself is correct. What confuses me is that different members retargeted the same original AMASS SFU 0007 walking motion separately, yet the expert data ends up different and greatly affects training results. Could you explain a bit more about how your individual retargeting processes differ?

[JinhaoGao]

Thank you for your response. Because we needed compatibility with another project's version, we adopted Isaaclab version 5.0.0 in this project to introduce viscous friction. Regarding the expert's animation, could you provide a suggestion on how to judge its quality? After all, visualization alone cannot visually convey the quality of different expert data, and it does indeed affect the training results.Furthermore, during the GMR redirection process, I think the differences might lie in the parameters of the ik table.Although the difference is imperceptible to the naked eye—neither the replay of the redirected movements nor the replay of the movements after switching to expert-level data shows any difference—it does indeed affect training effectiveness. Therefore, I would like to ask you for some good methods for judging the quality of expert data? Looking forward to your reply.