Accepted for publication in IEEE ICAR 2021
Description:
In this work we present two novel strategies for training Feedforward Artificial Neural Networks (ANN) to learn robot’s forward kinematics (3D position only).
Generally, ANN are used to diectly learn the mapping from actuation space to Cartesian space using a cost function that just minimizes the error between the predicted tip position and the measured (ground truth) one.
For control purposes, however, the velocity mapping is exploited. This information is neglected instandard ANN training.
The two novel loss functions include information about the differential reletionship between positions and velocities to incorporate knowledge on the velocity mapping too while training the model for the forward kinematics.
Our results show large improvements in model learning and control when the differential relationship is included.
Francesco Cursi 