Introduction Stroke is always associated with a difficult functional recovery process. A brain-computer interface (BCI) is a technology which provides a direct connection between the human brain and external devices. The primary aim of this study was to determine whether training with a BCI-controlled robot can improve functions in patients with subacute stroke. Methods Subacute stroke patients aged 32-68 years with a course of 2 weeks to 3 months were randomly assigned to the BCI group or to the sham group for a 4-week course. The primary outcome measures were Loewenstein Occupational Therapy Cognitive Assessment (LOCTA) and Fugl-Meyer Assessment for Lower Extremity (FMA-LE). Secondary outcome measures included Fugl-Meyer Assessment for Balance (FMA-B), Functional Ambulation Category (FAC), Modified Barthel Index (MBI), serum brain-derived neurotrophic factor (BDNF) levels and motor-evoked potential (MEP). Results A total of 28 patients completed the study. Both groups showed a significant increase in mean LOCTA (sham: P < 0.001, Cohen's d = - 2.972; BCI: P < 0.001, Cohen's d = - 4.266) and FMA-LE (sham: P < 0.001, Cohen's d = - 3.178; BCI: P < 0.001, Cohen's d = - 3.063) scores. The LOCTA scores in the BCI group were 14.89% higher than in the sham group (P = 0.049, Cohen's d = - 0.580). There were no significant differences between the two groups in terms of FMA-B (P = 0.363, Cohen's d = - 0.252), FAC (P = 0.363), or MBI (P = 0.493, Cohen's d = - 0.188) scores. The serum levels of BDNF were significantly higher within the BCI group (P < 0.001, Cohen's d = - 1.167), and the MEP latency decreased by 3.75% and 4.71% in the sham and BCI groups, respectively. Conclusion Training with a BCI-controlled robot combined with traditional physiotherapy promotes cognitive function recovery, and enhances motor functions of the lower extremity in patients with subacute stroke. These patients also showed increased secretion of BDNF.