In this paper the influence of varying the starting conditions on intensity-modulated proton therapy (IMPT) plans has been studied. In particular IMPT plans have been optimized based on four different starting conditions of initial beamlet fluences: (a) all beamlets with an initial constant weight, delivering a gradient from the proximal to the distal edge of the target (forward wedge approach); (b) beamlet weights reduced from the distal to the proximal aspect of the target such as to deliver a flat 'spread-out-Bragg-peak' (SOBP approach); (c) beamlet weights calculated to deliver a gradient from the distal (maximal dose) to the proximal edge (inverse wedge); (d) beamlet weights set universally to zero except the most distal one, for each given lateral direction (i.e. distal-edge-tracking, DET). An analysis of robustness to range errors has been performed by recalculating plans, assuming a systematic 3% error in CT values. Results showed that IMPT plans optimized with the forward wedge approach were very sensitive to range errors, since organs-at-risk (OAR) were spared by patching single-field lateral and distal fall-offs, the last ones being strongly sensitive to range errors. In addition a plan robust to range errors can be achieved by starting the optimization process in the case of low-dose constraints to OAR, with the initial flat SOBP approach, and with either the DET or the inverse wedge approaches, in the case of stringent dose-volume constraints to OAR. 'Starting condition-based optimization' as proposed here can therefore provide a tool to transparently 'steer' the optimization outcome to solutions more robust to uncertainties.