I am training a 4-DOF OpenManipulator-X arm in Isaac Lab for a pick-and-place task. I am encountering a persistent issue where the parallel prismatic gripper fingers (left and right) do not move symmetrically during policy training.
While the URDF mimic constraints work perfectly in the Isaac Sim Physics Inspector (manually moving one slider moves both fingers), they fail during RL training. The fingers often slide in the same direction, drift apart, or ignore the mimic constraint entirely when driven by the ActionManager via BinaryJointPositionActionCfg or JointPositionActionCfg.