Evaluation of a Model-free Approach to Object Manipulation in Robotics

Sammanfattning: Action Recognition is crucial for object manipulation in robotics. In recent years, Programming by Demonstration has been proposed as a way for a robot learning tasks from human demonstrations. Based on this concept, a model-free approach for object manipulation has recently been proposed in [1]. In this thesis, this model-free approach is evaluated for Action Recognition. In specific, the approach classifies actions by observing object-interaction changes from video. Image segmentation to videos presents various difficulties, such as motion blur, complex environment, Over- and Under- segmentation. This thesis investigates and simulates these image segmentation errors in a controllable manner. Based on the simulation, two different similarity measure methods are evaluated: The Substring Match (SSM) and Bhattacharyya Distance (B-Distance) method. The results show that the B-Distance method is more consistent and capable to classify actions with higher noise level compare to the SSM method. Further, we propose an action representation using kernel method. The evaluation shows that the novel representation improves Action Recognition rate significantly.