Neural network based estimator and controller for SLIP and TD-SLIP monopod robots

Abstract

Using spring loaded inverted pendulum models for legged locomotion, a wide range of applications can be developed for mobile robots. Spring loaded inverted pendulum model brings new challenges of solving the forward and inverse kinematic maps. Although exact solutions of SLIP model cannot be obtained analytically due to nonintegrability of stance dynamics, approximate analytical solutions are proposed to overcome these challenges in many of the previous studies. An alternative to approximate analytical solutions, neural network based forward and inverse kinematic map predictions can also be used. We used neural networks to design estimators and controllers, as forward and inverse kinematic map predictors. Required datasets are generated with existing simulations for spring loaded inverted pendulum models and datasets are constructed by including the determined inputs and outputs for the neural networks. Neural networks are designed by using different architectures and properties. Training results of estimators for predicting the goal state in the apex return map are reported for different configurations. Trained controllers are verified using the simulation and verified controllers are tested under the different run scenarios. By comparing all of the results, potential of the neural network based estimators and controllers is discussed.