Calibration issue?

Hello all,

We noticed some consistent errors when trying to grasp the cuboid, even after averaging over multiple observations to reduce noise in the objects pose. Looking at the picture below the grasp looks like it would be successful given the observed object and robot pose, but the object’s actual position is not aligned with the fingertips.

Since the detection of the cuboid looks good, this seems to be a small calibration issue in either the cameras or the robot. This was observed on roboch6, but similar errors were also seen on roboch1, 4, and 7.

Is this something you can look into fixing? Or should we deal with this level of error on our side?

Thanks for all your hard work!!

-Chip Schaff

1 Like

Thanks for reporting, I will take a look if we can improve the calibration.
There will always be a bit of an error, though, due to the mechanical properties of the robot: To get the angular positions of the joints, we are using encoders which are attached to the motors. From the motor position we compute the joint position using the known gear ration. This has some inaccuracy, as the actual joint position can differ a bit from the computed one due to backlash and slight stretching of the timing belts.

1 Like

Related to this, we may want to try running calibration adjustment on our side as well, using the camera images and motor position readings.

Could it be possible for you to release real camera intrinsic parameters and an algorithm that detects tip positions or robot poses in an image (if there’s anything like that…)

Best,
Takuma

Sorry that it took so long. There was indeed some error in the calibration, see Fix in Robot Calibration. At least on the robot I tested with, this significantly improved the match between real robot and visualisation.

Regarding @takumaynd’s question:
The camera calibration parameters are not affected by this issue. They are available during runtime in /etc/trifingerpro, see documentation. A copy of these files is added to the log files of each submission, there they are called camera{60,180,300}.yml.

Unfortunately, we don’t have any algorithm for estimating the robot pose via the cameras.

1 Like