In this paper, an FIM (Fitness to Ideal Model)and a DLen (Description Length) based evaluation approachhas been developed to measure the benefit of learning from experienceto improve the robustness of the robot’s behavior. Theexperience based mobile artificial cognitive system architectureis briefly described and adopted by a PR2 service robot withinthe EU-FP7 funded project RACE. The robot conducts typicaltasks of a waiter. Temporal and lasting obstacles and standardtable items, as shown in the demonstrations of “Deal-withobstacles”and “Clear-table-intelligently”, are being adoptedin this work to test the proposed evaluation metrics, validateit on a real PR2 robot system and present the evaluationresults. The relationship between the FIM and DLen has beenvalidated. This work proposes an effective approach to evaluatea cognitive service robot system which enhances its performanceby learning.