Inverse Kinematics Analysis of Cassie Robot using Radial Basis Function Networks

Inverse Kinematics of bipedal humanoid robots remains a challenging problem in the domain of robotics and computation, due to high order non-linearity and computation involved in Inverse Kinematics solutions. Also, there are many constraints involved with the various joint parameters which makes their analysis even more complex. Through this paper, we attempt to solve the Inverse Kinematics problem of a bipedal humanoid robot, Cassie, using Radial Basis Function (RBF) Networks. Our method can also be applied to other higher degrees of freedom serial manipulators. Our simulation analyses the results based on size of datasets, data distribution and network parameters. We have considered datasets of size $\sim$300k and $\sim$1 million, single and multiple hidden layers, equal and random data distribution, different number of neurons in layers and different training functions. We achieve our target of limiting the Mean Squared Error (MSE) calculated using the trained model below 0.1° for each joint angle, which is under acceptable limits for practical implementation. The configurations obtained from the RBF network are simulated and compared with the original input configuration. This is compared visually in MATLAB and the resulting pose of end-effector are similar for both the cases, complementing the performance that we get for the networks.