Background Closed-loop insulin delivery systems are expected to become a standard treatment for patients with type 1 diabetes. We aimed to assess whether the Diabeloop Generation 1 (DBLG1) hybrid closed-loop artificial pancreas system improved glucose control compared with sensor-assisted pump therapy. Methods In this multicentre, open-label, randomised, crossover trial, we recruited adults (aged ≥18 years) with at least a 2 year history of type 1 diabetes, who had been treated with external insulin pump therapy for at least 6 months, had glycated haemoglobin (HbA1c) of 10% or less (86 mmol/mol), and preserved hypoglycaemia awareness. After a 2-week run-in period, patients were randomly assigned (1:1) with a web-based system in randomly permuted blocks of two, to receive insulin via the hybrid closed-loop system (DBLG1; using a machine-learning-based algorithm) or sensor-assisted pump therapy over 12 weeks of free living, followed by an 8-week washout period and then the other intervention for 12 weeks. The primary outcome was the proportion of time that the sensor glucose concentration was within the target range (3·9–10·0 mmol/L) during the 12 week study period. Efficacy analyses were done in the modified intention-to-treat population, which included all randomly assigned patients who completed both 12 week treatment periods. Safety analyses were done in all patients who were exposed to either of the two treatments at least once during the study. This trial is registered with ClinicalTrials.gov, number NCT02987556. Findings Between March 3, 2017, and June 19, 2017, 71 patients were screened, and 68 eligible patients were randomly assigned to the DBLG1 group (n=33) or the sensor-assisted pump therapy group (n=35), of whom five dropped out in the washout period (n=1 pregnancy; n=4 withdrew consent). 63 patients completed both 12 week treatment periods and were included in the modified intention-to-treat analysis. The proportion of time that the glucose concentration was within the target range was significantly higher in the DBLG1 group (68·5% [SD 9·4] than the sensor-assisted pump group (59·4% [10·2]; mean difference 9·2% [95% CI 6·4 to 11·9]; p<0·0001). Five severe hypoglycaemic episodes occurred in the DBLG1 group and three episodes occurred in the sensor-assisted pump therapy group, which were associated with hardware malfunctions or human error. Interpretation The DBLG1 system improves glucose control compared with sensor-assisted insulin pumps. This finding supports the use of closed-loop technology combined with appropriate health care organisation in adults with type 1 diabetes.