Ricerc@Sapienza

Data-driven modeling of dynamical systems gathers attention in several applications; in conjunction with model predictive control, novel different identification techniques that merge machine learning and optimization are presented and compared with the purpose of reducing seismic response of frame structures and minimize control effort. Performance of neural network-, random forest- and regression tree-based identification algorithms in producing reliable models exploiting historical data coming from a real structure is shown.

A redundant robotic system must execute a task in a workspace populated by obstacles whose motion is unknown in advance. For this problem setting, we present a sensor-based planner that uses Model Predictive Control (MPC) to generate motion commands for the robot. We also propose a real-time implementation of the planner based on ACADO, an open source toolkit for solving general nonlinear MPC problems. The effectiveness of the proposed algorithm is shown through simulations and experiments carried out on a UR10 manipulator.

This paper presents a Model Predictive Control (MPC) scheme capable of generating a 3D gait for a humanoid robot. The proposed method starts from an assigned sequence of footsteps and generates online the trajectory of both the Zero Moment Point and Center of Mass. Starting from the moment balance (neglecting rotations) we derive a model characterizing all 3D trajectories that satisfy a linear differential equation along all three axes.