Questions before direct GO-1 deployment on purchased G1 robots (DoF mismatch, head/waist defaults, gripper type, camera placement)

[GuaguaguaLiu]

Hi GO-1 team,

Thanks a lot for open-sourcing this great work.
We are trying to directly deploy the released GO-1 model on our purchased AgiBot G1 robots for some initial task trials, and we'd really appreciate clarification on several practical deployment questions:

1. Training data vs model output dimensions

• From the dataset/docs/issues, we understand raw data can include head/waist/base-related signals (e.g., 20/22 dims in some converted settings), while released GO-1 checkpoint outputs 16-dim actions.
• Could you clarify the exact action/state definition used to train the released GO-1 checkpoint?
• Why is there this mismatch between dataset-available dimensions and released model output dimensions?

2. If head/waist DoFs are not modeled in GO-1 action output

• During real deployment, should head/waist be set to fixed defaults to match pretraining distribution?
• If yes, could you share recommended default values/ranges (or a calibration procedure) for head/waist so the camera viewpoint is consistent with training?

3. Gripper variants on G1

• We have two purchased G1 units with different grippers:
  • OmniPicker gripper
  • older black gripper (not sure of official name)
• For direct zero-shot deployment with released GO-1, which gripper type is more aligned with your pretraining distribution?
• Do you recommend separate gripper mapping/calibration for these two variants?

4. Wrist camera placement requirements

• Our wrist camera positions were adjusted after purchase.
• How sensitive is GO-1 to wrist camera extrinsics/placement changes?
• Is there any recommended geometric standard, tolerance range, or alignment checklist to keep inference in-distribution?

If there are any reference configs/checklists for direct G1 deployment (especially for head/waist defaults + camera alignment + gripper mapping), we'd be very grateful.

Thanks again!