Background. More than two-thirds of stroke patients have arm motor impairments and function deficits on hospital admission, leading to diminished quality of life and reduced social participation. Robot-assisted training (RAT) is a promising rehabilitation program for upper extremity while its effect is still controversial due to heterogeneity in clinical trials. We performed a systematic review and meta-analysis to compare robot-assisted training (RAT) versus therapist-mediated training (TMT) for arm rehabilitation after stroke. Methods. We searched the following electronic databases: MEDLINE, EMBASE, Cochrane EBM Reviews, and Physiotherapy Evidence Database (PEDro). Studies of moderate or high methodological quality (PEDro score ≥4) were included and analyzed. We assessed the effects of RAT versus TMT for arm rehabilitation after stroke with testing the noninferiority of RAT. A small effect size of -2 score for mean difference in Fugl-Meyer Assessment of the Upper Extremity (FMA-UE) and Cohen’s d = -0.2 for standardized mean difference (SMD) were set as noninferiority margin. Results. Thirty-five trials with 2241 participants met inclusion criteria. The effect size for arm motor impairment, capacity, activities of daily living, and social participation were 0.763 (WMD, 95% CI: 0.404 to 1.123), 0.109 (SMD, 95% CI: -0.066 to 0.284), 0.049 (SMD, 95% CI: -0.055 to 0.17), and -0.061 (SMD, 95% CI: -0.196 to 0.075), respectively. Conclusion. This systematic review and meta-analysis demonstrated that robot-assisted training was slightly superior in motor impairment recovery and noninferior to therapist-mediated training in improving arm capacity, activities of daily living, and social participation, which supported the use of RAT in clinical practice. Copyright © 2020 Zejian Chen et al.