Grape cluster compactness affects berry ripening homogeneity, fungal disease incidence, grape composition and wine quality. Therefore, assessing cluster compactness is crucial for sorting wine grapes for the wine industry. Nowadays, cluster compactness assessing methodology is based either on visual inspection performed by trained evaluators (OIV method) or on morphological features of clusters. The goal of this work was to develop an innovative and automated, non-destructive method to assess cluster compactness, based on computer vision on-the-go, under field conditions. RGB images were acquired on October 2016, before harvest, in a Tempranillo (Vitis vinifera L.) commercial vineyard. An all-terrain vehicle (ATV) was modified with equipment to autonomously capture images of 95 clusters, one week prior to harvest. Tempranillo vineyard located in La Rioja (Spain) was trained to a vertically shoot-positioned (VSP) trellis system and partially defoliated before veraison. Image acquisition was conducted using artificial illumination at night-time while the ATV moved at 5 km/h. The next day, all photographed clusters were collected, and their compactness rating was assigned by a panel of ten trained experts following the OIV 204 standard code. A combination of different machine learning and computer vision techniques was used to determine the compactness from the acquired images, obtaining a R2 of 0.71 and RMSE of 1.208 calculated by the leave-one-out cross-validation (LOOCV), using the mean of the expert’s evaluation as the reference value. These results show a strong correlation between the algorithm estimation and the average rating of a group of trained evaluators, suggesting that an automatic system can be applied to estimate cluster compactness in vineyards as an efficient alternative to traditional visual methods.