Gravity Pivoting from Motor Current

Calibration-Free Angle Estimation and the Sensing Budget It Requires

Abstract

Gravitational pivoting lets a parallel-jaw gripper reorient a grasped object by loosening its grip so that gravity rotates the object about the grasp axis. Published pivoting controllers observe the object with tactile sensors or cameras. We ask whether the motor current of the gripper’s wrist actuator is enough. Under quasi-static conditions the wrist carries the gravity torque m g r cos(theta + psi), so a slow downhill sweep of the wrist angle psi while the object is clamped, followed by a linear least-squares fit, recovers the direction theta of the object’s centre of mass. The estimate needs no object mass, grasp offset or motor torque constant, and a sign-of-velocity regressor removes gearbox Coulomb friction. In MuJoCo simulation with a soft-finger contact model and 100 randomized objects, the estimator located the centre-of-mass direction to 0.009 rad RMS with ideal sensing. Through a modelled current-sensing chain it stayed at 0.0093-0.0094 rad for gearbox friction up to 0.1 N m, while a single static current reading degraded from 0.088 to 1.39 rad. A linearized error budget in ADC resolution and current noise matched simulation to within a factor of 0.87-1.12 whenever the noise exceeded half a quantization step, and failed by up to 26.7x otherwise; adding noise improved an 8-bit sensor 12-fold. In closed loop, a controller using only motor current pivoted 40 randomized objects to target angles with 0.031 rad RMS error (90% within 0.05 rad), against 0.107 rad for an open-loop schedule given the true object parameters and 0.413 rad for one given nominal parameters. We also report the limits we found: torque observes the centre-of-mass direction rather than the object’s geometric axis, a wrist motion made while the grip is loose ratchets the object downward, and the controller needs about 16.6 probes per pivot. All results are from simulation.