This paper studies the use of predictive in-cycle close-loop combustion control to reduce the stochastic cyclic variations of diesel combustion. The combustion metrics that fully define the pressure trace with a pilot-main injection i.e. pilot and main start of combustion, burned pilot mass, and engine load are used as the set-point reference. These metrics are in-cycle predicted by calibrated models as functions of the current cylinder state, estimated by in-cylinder pressure measurements. The proposed approach uses four individual controllers for the set-point error minimization, which respectively regulate the injection’s timing and duration of the pilot-main injection. The controllers are implemented in a FPGA and tested in a Scania D13 engine. The steady-state error reduction, disturbance rejection and transient response are discussed. The results confirm the error reduction in both, cycle-to-cycle and cylinder-to-cylinder variations. The error dispersion, measured by the 95% confidence interval, was reduced between 25% and 75% for all the controlled parameters. By on-line adaptation, the controllers are robust against model uncertainties and fuel types.