The saccade generator updates memorized target representations for saccades during eye and head movements. Here, we tested if proprioceptive feedback from the arm can also update hand-held object locations for saccades, and what intrinsic coordinate system(s) are used in this transformation. We measured radial saccades beginning from a central light-emitting diode to sixteen target locations arranged peripherally in eight directions and two eccentricities on a horizontal plane in front of subjects. Target locations were either indicated by 1) a visual flash, 2) by subject actively moving the hand-held central target to a peripheral location, 3) by the experimenter passively moving the subject’s hand, 4) through a combination of the above proprioceptive and visual stimuli. Saccade direction was relatively accurate, but subjects showed task-dependent systematic overshoots and variable errors in radial amplitude. Visually-guided saccades showed the smallest overshoot, followed by saccades guided by both vision and proprioception, while proprioceptively-guided saccades showed the largest overshoot. In most tasks, the overall distribution of saccade endpoints was shifted and expanded in a gaze- or head-centered cardinal coordinate system. However, the active proprioception task produced a tilted pattern of errors, apparently weighted toward a limb-centered coordinate system. This suggests the saccade generator receives an efference copy of the arm movement command but fails to compensate for the arm’s inertia-related directional anisotropy. Thus, the saccade system is able to transform hand-centered somatosensory signals into oculomotor coordinates and combine somatosensory signals with visual inputs, but it appears to have a poorly calibrated internal model of limb properties.