Uncertainty Quantification of Robot Inverse Kinematics with Neural Networks

In motion control of the robotic arm, it is computationally quite complicated to find an accurate and reliable solution for inverse kinematics. Hence we develop a Neural Network(NN) model that estimates the IK values as well as the uncertainty of each estimation. Even though NNs have proven to give highly accurate predictions and have reached many milestones, it also tends to produce over-confident predictions at certain times. Therefore, it becomes imperative for the model to know its unreliability when facing unknown domains.

An uncertainty assessment is essential because we can determine the shortcomings, and the model can be calibrated for a better quality of predictions. There is quite a few robotic research showing that estimating uncertainty is profitable. However, there is very little research comparative study on which of the uncertainty technique performs better. Another core challenge we will be working on here is the out-of-distribution(OOD) detection problem, where we would check if the model can produce a higher uncertainty value for anomalous input or even co-variate shifts concerning their training data.

The main goal of the thesis is to estimate the uncertainty in the Inverse kinematics prediction model. We would initially analyse different Uncertainty quantification techniques such as Deep Ensembles, MC Dropout and MC DropConnect on the Inverse Kinematic prediction neural network. This evaluation will also include tuning hyper-parameters and conducting various experiments to understand the data we have and the model better. We should also be able to infer the optimal technique for this robotic application. Further on, we also analyse uncertainty exclusively in the case of OOD data. As a part of the OOD experiment, we apply various types of split on the dataset and train the NN with partial data, thus proving higher uncertainty in OOD areas while testing. Finally, being able to accurately estimate the uncertainty of the deep learning model for robotic applications.

In der Regel sind die Vorträge Teil von Lehrveranstaltungsreihen der Universität Bremen und nicht frei zugänglich. Bei Interesse wird um Rücksprache mit dem Sekretariat unter sek-ric(at)dfki.de gebeten.

last updated 30.07.2019
to top