The torsional oscillations of the driveline represent a well-known issue in electric powertrains. In order to dampen such oscillations, active damping control is usually employed to suitably modulate the requested torque. However, on the one hand, the modelling inaccuracies due to the presence of nonlinear elements in the driveline system might lead to an ineffective model-based calibration. On the other hand, model-free iterative data-driven calibration overcomes the above problem at the cost of possible safety issues. In fact, unsafe controllers might be tested producing potential system failures. This paper presents a safe model-free calibration framework based on Bayesian optimization, aiming at optimizing the latent objective function while minimizing the risk - in a probabilistic manner - of critical explorations. Results on a full-fledged vehicle simulator shows that the performance is improved as compared to model-based calibration, while at the same time the experimental effort is concentrated in safe operating regions.