In this paper, an FIM (Fitness to Ideal Model) and a DLen (Description Length) based evaluation approach has been developed to measure the benefit of learning from experience to improve the robustness of the robot's behavior. The experience based mobile artificial cognitive system architecture is briefly described and adopted by a PR2 service robot within the EU-FP7 funded project RACE. The robot conducts typical tasks of a waiter. Temporal and lasting obstacles and standard table items, as shown in the demonstrations of 'Deal-with-obstacles' and 'Clear-table-intelligently', are being adopted in this work to test the proposed evaluation metrics, validate it on a real PR2 robot system and present the evaluation results. The relationship between the FIM and DLen has been validated. This work proposes an effective approach to evaluate a cognitive service robot system which enhances its performance by learning.