Job Description
I am building a system to press a push-to-open button using a Kinova Gen3.
The inverse kinematics is fully working. If I provide an (x, y, z, quaternion) pose in base_link, the arm reaches the target without any issue.
The perception side is also working. I am using a custom YOLO26n-seg model for segmentation, and it reliably detects the button. The camera is an Intel RealSense mounted on the wrist. Depth alignment and centroid extraction from the segmentation mask are functioning correctly.
The problem is in the transformation pipeline. After computing the centroid, converting it to a 3D point in the camera frame, and transforming it to base_link, the resulting target pose dis off (not where the button actually is). This appears to be a TF or projection-related bug.
Possible causes include incorrect camera intrinsics usage, improper depth scaling, misunderstanding of the optical frame convention, an incorrect static transform between the camera and the wrist, or an incorre...
The inverse kinematics is fully working. If I provide an (x, y, z, quaternion) pose in base_link, the arm reaches the target without any issue.
The perception side is also working. I am using a custom YOLO26n-seg model for segmentation, and it reliably detects the button. The camera is an Intel RealSense mounted on the wrist. Depth alignment and centroid extraction from the segmentation mask are functioning correctly.
The problem is in the transformation pipeline. After computing the centroid, converting it to a 3D point in the camera frame, and transforming it to base_link, the resulting target pose dis off (not where the button actually is). This appears to be a TF or projection-related bug.
Possible causes include incorrect camera intrinsics usage, improper depth scaling, misunderstanding of the optical frame convention, an incorrect static transform between the camera and the wrist, or an incorre...
Apply for this Position
Ready to join Confidential? Click the button below to submit your application.
Submit Application