How were the camera poses in calibration/ generated?
Hi, thanks for releasing the per-frame camera poses for flex-pi/robotwin_3d, they are very useful.
The npz files record robotwin_commit 266f3aad… and an asset_archive_sha256. Were the poses produced by replaying each episode in the RoboTwin simulator, or computed from forward kinematics of the recorded joint states plus a camera mount?
If they came from a replay, would you be willing to share the generation script, or the per-episode seeds and scene configuration? With those we could also export object poses and segmentation for the same episodes.
We also see the left-wrist camera disagree with the rendered depth by a few millimetres on some episodes (for example episode 9833), while the head and right wrist agree to below 1 mm. Is that expected?
Thanks!