Uncertainty Quantification of Robot Inverse Kinematics with Neural Networks

Application of Inverse Kinematics(IK) prediction is on multi-folds from robotics, computer graphics, and video games, while recently being used in protein chain prediction and rehabilitation physiology. Although there are multiple methods to solve and predict IK, they are very less likely to generate the uncertainty of the model as well.

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 Deep Neural Network(DNN) model that estimates the IK values as well as the uncertainty of each estimation. Even though DNNs have proven to give highly accurate predictions and have reached many milestones, it also has a tendency to produce over-confident predictions at certain times. Hence, it becomes imperative for the model to know its unreliability when facing unknown domains.

An uncertainty assessment is important because we can determine the shortcomings and the model can be "calibrated" for the better quality of predictions. Another core challenge we will be working on here is the out-of-distribution(OOD) detection problem, where we would check if the model is able to produce higher uncertainty value for anomalous input or even covariate shifts with respect to their training data.

The main goal of the thesis is to estimate the uncertainty in Inverse kinematics prediction models. We would initially analyze different Uncertainty quantification techniques such as Deep Ensembles, MC Dropout, MC DropConnect on the Inverse Kinematics models. This evaluation will also include tuning of hyper-parameters in this case, the degree of freedom and the kinematic configurations such as joint angles and corresponding end-effector values, etc. Eventually, being able to accurately estimate the uncertainty of the deep learning model for varying conditions.