GO-1 action space

[MSheng-Lee]

@ModiShi
Thank you for your great job!
Could you clarify whether GO-1 was trained using absolute or delta positions for actions? Also, regarding the action ordering left_arm_joints: [0:7], left_arm_gripper: [7], right_arm_joints: [8:15], right_arm_gripper: [15] - should these be directly concatenated in this exact order? I want to ensure I'm interpreting the dataset_stats.json format correctly for deployment.

Does the model use any additional action dimensions beyond joints/grippers, such as end (orientation / position), waist (effort/position/velocity)?

[zbzyjya]

我感觉是的 我拿训练集再次训练时 以绝对关节角作（也就是你所描述的16维度）为输入的时候 loss非常小

[MSheng-Lee]

我感觉是的 我拿训练集再次训练时 以绝对关节角作（也就是你所描述的16维度）为输入的时候 loss非常小

@zbzyjya
感谢您的澄清！关于训练数据集中动作空间定义的另一个问题：在您的动作表示中，关节位置值是使用 /state/joint/position 还是 /action/joint/position 呢？

[zbzyjya]

我感觉是的 我拿训练集再次训练时 以绝对关节角作（也就是你所描述的16维度）为输入的时候 loss非常小

@zbzyjya 感谢您的澄清！关于训练数据集中动作空间定义的另一个问题：在您的动作表示中，关节位置值是使用 还是 呢？/state/joint/position``/action/joint/position

在训练配置文件中分别需要填写action_dim和state_dim 我的理解是原模型在训练时这两个参数都是16 然后训练的输入/action/joint/position对应action /state/joint/position对应state

这是我观察所得 若有说的不对的地方 请指正

[ModiShi]

We use /state/joint/position as the action.

[MSheng-Lee]

We use /state/joint/position as the action.

@ModiShi

I noticed an inconsistency in the normalization statistics between the published dataset_stats.json file (https://huggingface.co/agibot-world/GO-1/blob/main/dataset_stats.json) and the statistics.py implementation (https://github.com/AgibotTech/Genie-Envisioner/blob/819c27a96e297e271fddc5f248dbd8860a644758/data/utils/statistics.py#L11). Could you clarify which should be used as the authoritative source for normalization during fine-tuning?

[ModiShi]

@MSheng-Lee We used https://huggingface.co/agibot-world/GO-1/blob/main/dataset_stats.json for GO-1. The pretraining data may not be fully consistent between GO-1 and Genie-Envisioner, which could account for this difference.