Clarification on Plug Orientation Relative to Gripper Across Evaluation Trials

Competitions AI for Industry Challenge
1

Hi,

I have a question regarding the plug’s starting pose relative to the gripper during evaluation.

Looking at the sample_config.yaml, I see that the cable is spawned with a specific gripper_offset(x, y, z) and orientation (roll, pitch, yaw) relative to the end-effector. In the sample config, these values appear consistent across all three trials:

pose:

gripper_offset:

x: 0.0

y: 0.015385

z: 0.04245   # varies slightly per trial

roll: 0.4432

pitch: -0.4838

yaw: 1.3303

My questions:

  1. Is the plug-to-gripper orientation (roll/pitch/yaw) guaranteed to be the same fixed values across all evaluation trials? Or could the plug be grasped at a different/randomized orientation relative to the gripper?

  2. Can we treat the plug tip position as a known, static offset from gripper/tcp (since ground-truth TFs for the cable are not available during evaluation)?

  3. Will the slight variation in the z-component of gripper_offset seen in the sample config also appear in evaluation, or will this be fixed?

Understanding this will help determine whether our policy needs to solve for plug localization (via perception) in addition to port localization, or whether we can rely on the plug pose being a known constant relative to the gripper.

Thank you!

1 Like
2

@Rocky_Shao thanks for raising this question.

We updated our docs to clarify this point Clarify variation in grasp poses by Yadunund · Pull Request #412 · intrinsic-dev/aic · GitHub.

You can expect the poses to be relatively consistent for each plug type.

2 Likes