Neural Network-Based Reaction Estimator for Walking Robots

Abstract

The paper focuses on one of the problems associated with controlling walking robots using model-based control methods: the presence of explicit mechanical constraints in the robot's dynamical model. The paper proposes the use of a neural network-based estimator for the reaction forces allowing the simplification of the mentioned control problem. Two architectures of the neural network are proposed, both are based on fully connected layers with rectified linear unit activation functions. The generation of the training dataset for the network is discussed. It is shown that the trained network is capable of accurately predicting the values of normal reactions using the values of generalized coordinates and velocities, as well as control actions as inputs. The predictions demonstrate a constant offset, which we refer to as the static prediction error. The paper presents some possible use cases for the designed reaction estimator.