It is well known that algorithms for measuring available path capacity at the IP layer by active probing tend to be more accurate when using long probe trains. This is due to the bias errors introduced by short probe trains. However, long trains have other disadvantages, such as delay, overhead, and increased risk for packet loss, especially for high probe rates. Router queue sizes also limit the maximum probe train length. In this paper, a robust method is proposed for reducing measurement bias caused by short probe trains. The method is robust in the sense that it gracefully handles lost packets and outliers, such as caused by non-Poisson cross traffic. The proposed method has been combined with the BART estimator for measuring available path capacity, and successfully tested in a multi-hop local IP network for industrial use. The principle should be generally applicable to estimators of available path capacity based on active probe trains.