Position-Based Visual Servoing

A visual servo method that first estimates the target's 3D pose from the image, then drives the robot to reduce the pose error.

Visual servoing means feeding camera images directly into the control loop. The classic 2006 tutorial by François Chaumette and Seth Hutchinson splits it into two families: PBVS and image-based visual servoing (IBVS). PBVS first uses the camera's calibrated intrinsics and a 3D model of the object to estimate the camera's pose relative to the target from the image, then defines an error e as the difference from the desired pose — usually written as a translation t plus an angle-axis rotation θu. The camera velocity command is v = −λL⁺e, where λ is a gain and L⁺ is the pseudoinverse of the interaction matrix (how the error changes with camera velocity). The advantage is that the error lives in 3D space, so the camera's path is intuitive and close to a straight line. The drawback is heavy reliance on calibration and pose estimation accuracy: if the estimate is off, the robot simply stops at the wrong pose, and the target can drift out of view during motion. IBVS instead compares image feature points directly, which tolerates calibration error better.

ExampleA wrist camera (eye-in-hand) looks at a part on the table. Each frame, pose estimation computes the part's 6D pose relative to the camera, subtracts it from the desired pre-grasp camera pose to get the error, and keeps correcting the end-effector velocity until the error converges to zero before the gripper closes.

Also called

PBVS, Pose-Based Visual Servoing, 3D Visual Servoing

Related

Visual Servoing · Image-Based Visual Servoing · 6D Object Pose Estimation · Camera Calibration · Hand-Eye Calibration · Eye-in-Hand

Sources

Chaumette & Hutchinson, Visual Servo Control Part I: Basic Approaches (IEEE RAM, 2006)
Wikipedia: Visual servoing