Application of neural networks in inverse dynamics based contact force estimation

In the majority of robotic applications, including manipulation and human-robot interaction, contact force needs to be monitored and controlled. Compliance controllers demand high precision force measurement that can be delivered by commercial force/torque sensors. However, these sensors are expensive, rather bulky and vulnerable to impact forces. A common solution to this dilemma is the use of force observers, which estimate external forces using full knowledge about system dynamics. However, some robotic systems have complicated dynamics that may or may not be known entirely and precisely. In these situations the implementation of dynamic observers would not result in accurate force estimation. This paper proposes the use of neural networks in an inverse dynamics based force observation without the need for complete determination of system dynamics. We also show that for slow operations on soft environments, the observer estimates external forces without acceleration input