Equation (23) is printed with -\gamma\I_6; as a Tikhonov regularization it adds \gamma\I_6, as here. In \Um_g, the rotations are multiplied by a lever of 10 cm (the displacement they produce at 10 cm from the gripper), so that a single \gamma weights translations and rotations alike; the control law (24) itself does not depend on this scaling. The paper does not give its values of \alpha, B, and \gamma; the defaults here are \alpha=0.15, B=24, \gamma=0.2\,\mathrm{cm}^2.
Every action is limited to 0.3 cm and 2.5ยฐ (the robots' step); with the small gain, the actions change smoothly from one iteration to the next. The exploration moves each gripper by 0.5 cm or 4ยฐ along each axis and back: more than a control step, so that the responses stand out from the sampling noise of the tracked points (about the size of a pixel); the buffer of 24 keeps them during the first 12 actions of each gripper. Targets are produced by the same ARAP simulator, by moving the grippers from the rest shape, so they are reachable; they are deformed about 1.35 times more than the targets of pages 1 and 2. The camera samples the top surface on a fixed pixel grid of about 1.1 cm with 0.5 mm of noise; the mapping uses the parameters of the reference code (\Ii=4, \Iu=5, \Imax=30, steps of 4) and the anchored tracking of page 2.