In multiple sclerosis (MS) exercise improves upper limb functions, but it is unclear what training types are more effective. This study compares robot-assisted training based on haptic or sensorimotor exercise. 41clinically definite MS subjects with upper limb impairment were randomised into two groups: (i) Haptic and (ii) Sensorimotor. Subjects in the Haptic performed a robot-assisted training protocol designed to counteract incoordination and weakness. The task -interaction with a virtual mass-spring system against a resistive load- requires coordination skills. Task difficulty and magnitude of resistive load were automatically adjusted to the individual impairment. Subjects in the Sensorimotor performed reaching movements under visual control; the robot generated no forces. Both groups underwent eight training sessions (40 min/session, 2 sessions/week). Treatment outcome were 9HPT and ARAT scores. The average 9HPT score decreased from 74±9 s to 61±8 s for the Haptic and from 49±6 s to 44±6 s. We found a significant Treatment (p=.0453) and Time differences (p=.005), but no significant Treatment×Time interactions although we found that the absolute change was only significant in the Haptic group (p=.011). We observed no significant changes in the ARAT score. Participants tolerated treatments well with a low drop-out rate. In the subjects evaluated at after 12 week (11 subject in sensory-motor and 17 in haptic group) no retention of the effect was found. Task oriented training may improve upper limb function in persons with MS especially in prevalent pyramidal impaired subjects without maintain the effects after three months. NCT02711566 (clinicaltrial.gov). Copyright © 2020 Elsevier B.V. All rights reserved.