Vision based object grasping and manipulation in robotics require accurate estimation of object's 6D pose. The 6D pose estimation has received significant attention in computer vision community and multiple datasets and evaluation metrics have been proposed. However, the existing metrics measure how well two geometrical surfaces are aligned - ground truth vs. estimated pose - which does not directly measure how well a robot can perform the task with the given estimate. In this work we propose a probabilistic metric that directly measures success in robotic tasks. The evaluation metric is based on non-parametric probability density that is estimated from samples of a real physical setup. During the pose evaluation stage the physical setup is not needed. The evaluation metric is validated in controlled experiments and a new pose estimation dataset of industrial parts is introduced. The experimental results with the parts confirm that the proposed evaluation metric better reflects the true performance in robotics than the existing metrics.