Real-Time Implementation of State-Feedback Learning Framework for Bipedal Locomotion Over Undulating Surfaces

This article reports a unique state-feedback learning framework for the locomotion of a biped robot over undulating surfaces. The proposed framework has two main components: 1) Gait model: reference gait trajectory generation, and 2) state-feedback mechanism: integrated mechatronic setup of force sensitive resistor (FSR)-inertial measurement units (IMUs) sensor. The Gait model incorporates the feed-forward neural network-enabled universal activation function trained on the human gait dataset. A novel loss function that integrates the conventional with standard error is designed, which allows the training algorithm to fit low variance data-point precisely in trajectory to ensure physical constraints. Sensitivity analysis using the perturbation manifold-based influence measure is examined to quantify the impact of input uncertainty on model performance. The proposed mechatronics setup feeds the environmental undulating surface information to the Gait model. For this, Quaterion-EKF using Rodrigues parameter with mounted case compensation via Earth's gravity is designed to filter the IMUs readings. Moreover, Gaussian process regressor (GPR) model via Hamiltonian Monte Carlo is developed to map FSR's resistance to contact force. Here, the GPR model activates the feet IMUs that feed the ground slope to the Gait model. Finally, the validation of the proposed framework is presented for biped robot walking.