Background Robotic assisted movement has become an accepted method of treating the moderately-to-mildly impaired upper limb after stroke. Objective To determine whether, during the subacute phase of recovery, a novel type of robotic assisted training reduces moderate-to-severe impairment in the upper limb beyond that resulting from spontaneous recovery and prescribed outpatient therapy. Methods A prospective, randomized, double-blinded, placebo-controlled, semi-crossover study of 83 participants. Over 6- to 9-weeks, participants received 18, 30-min training sessions of the hand and wrist. The test intervention consisted of assisted motion, biofeedback, and antagonist muscle vibration delivered by a robotic device. Test Group participants received the test intervention, and Control Group participants received a placebo intervention designed to have no effect. Subsequently, Control Group participants crossed over to receive the test intervention. Results At enrollment, the average age (±SD) of participants was 57.0 ± 12.8 year and weeks since stroke was 11.6 ± 5.4. The average Fugl-Meyer baseline score of Test Group participants was 20.9, increasing by 10.8 with training, and in Control Group participants was 23.7 increasing by 6.4 with training, representing a significant difference (4.4) in change scores ( P = .01). During the crossover phase, Control Group participants showed a significant increase in FMA-UL score (i.e., 4.7 ± 6.7 points, P = .003) as well as in other, more specific measures of impairment. Conclusions Robotic impairment-oriented training, as used in this study, can significantly enhance recovery during the subacute phase of recovery. Spontaneous recovery and prescribed outpatient therapy during this phase do not fully exploit the potential for remediating moderate-to-severe upper limb impairment. ClinicalTrials.gov Registry: NCT00609115–Subacute stroke rehabilitation with AMES