Discriminative Predictive Analysis for Goal Prediction

Goal prediction has always been of interest for researchers. With the advent of robots in human life and humans working so closely with them, it is of paramount importance that goal prediction be looked at more closely to improve the human-robot interaction and prevent industrial accidents. The work that has been done in this field has been generative in nature. This thesis looks at discriminative goal prediction where it predicts the final goal of a robotic arm given its partial trajectory. Data for the experiment was collected from human teleoperation using Baxter robot and Microsoft Kinect. The features for learning are extracted from the partial trajectory. A logit model is used to fit the training data and predict from the test data. Both accuracy and log loss are used as evaluation criteria to see how well the model performs. The results verify the effectiveness of the discriminative goal prediction.