The ArmAssist is a simple low-cost robotic system for upper limb motor training that combines known benefits of repetitive task-oriented training, greater intensity of practice, and less dependence on therapist assistance. The aim of this preliminary study was to compare the efficacy of ArmAssist (AA) robotic training against matched conventional arm training in subacute stroke subjects with moderate-to-severe upper limb impairment. Twenty-six subjects were enrolled within 3 months of stroke and randomly assigned to the AA group or Control group (n = 13 each). Both groups were trained 5 days per week for 3 weeks. The primary outcome measure was Fugl-Meyer Assessment-Upper Extremity (FMA-UE) motor score, and the secondary outcomes were Wolf Motor Function Test-Functional Ability Scale (WMFT-FAS) and Barthel index (BI). The AA group, in comparison to the Control group, showed significantly greater increases in FMA-UE score (18.0 ± 9.4 versus 7.5 ± 5.5, p = 0.002) and WMFT-FAS score (14.1 ± 7.9 versus 6.7 ± 7.8, p = 0.025) after 3 weeks of treatment, whereas the increase in BI was not significant (21.2 ± 24.8 versus 13.1 ± 10.7, p = 0.292). There were no adverse events. We conclude that arm training using the AA robotic device is safe and able to reduce motor deficits more effectively than matched conventional arm training in subacute phase of stroke. The study has been registered at the ClinicalTrials.gov, ID: NCT02729649.