Visual Servoing & Precision Assembly (2D)

Visual Servoing Control

Idle — press Start

Part position offset (calibration error) 0.35

Shifts the true hole position without telling the controller — IBVS finds it purely from image feature error, no re-calibration needed.

Control gain (λ) 0.9

Feature error

–

Iteration

–

EE height

–

Insertions

How it works

The wrist-mounted camera is eye-in-hand: it looks straight down the same axis the gripper approaches along, so whatever it sees is exactly what the peg will meet. Each control tick it projects the four corners of the hole into the image plane and compares them against the desired corner positions — the pattern the corners would form if the end-effector were perfectly centred above the hole at the current height.

e = s_desired − s_current      (image feature error)
v = −λ · L⁺(s,Z) · e            (IBVS velocity command)
Z decreases ⇒ same offset → larger e (interaction matrix ∝ 1/Z)

Part position offset — the hole's real position never reaches the controller directly; only the image error does. Increase it and IBVS still converges, because it is closing the loop on pixels, not on a pre-programmed coordinate.
Control gain (λ) — how aggressively the velocity command follows the current feature error; too high overshoots and oscillates, too low converges slowly.
Feature error — the graph shows this shrinking iteration by iteration as the controller re-measures and re-corrects, the core IBVS convergence behaviour.

Real-world relevance: this closed image-based loop is what lets assembly robots tolerate fixture drift, thermal expansion and camera-to-gripper calibration error that would defeat an open-loop, coordinates-only pick-and-place program.

End-effector camera (top-down) Hole (true position) Peg / desired centre