The performance of Model-Predictive Control is fundamentally limited in tasks involving contact with the environment. In contrast, force control techniques are well-established in robotics, but they inherently lack planning capabilities. During my PhD, I proposed explored different ways to explain and bridge this critical gap. I developed, tested and evaluated novel MPC formulations that allow to systematically include measured efforts into modern optimal control.
The first idea I investigated during my PhD was to model the actuation dynamics in order to include joint torque measurement feedback [1]. This formulation has the avantage of not requiring additional sensors on torque-controlled robots while increasing the force control performance in dynamic contact tasks w.r.t. the standard MPC formulation.
Another idea relied on direct force modeling as a visco-elastic phenomena. This led to a powerful formulation that showed state-of-the-art performance in a wide range of tasks including force/motion tracking, multi-contactand and hard nonlinear constraints [2]. Crucially, this work demonstrates experimentally that MPC and force control can combine within a single control architecture without impeding any of their individual benefits, and outperforming each of them taken individually.
I also investigated another solution in collaboration with Armand Jordana, that relied on online estimation of the contact force modeling error in a disturbance-observer-like fashion [3]. This simple formulation allowed surprisingly good performance.
Formulating force tasks in arbiratry frames also required to derive contact dynamics derivatives in moving coordinates [4]. This formulation was added to the Crocoddyl optimal control library.
The force feedback MPC formulation has also been used in combination with higher level planning to achieve challenging industrial tasks such as deburring [5].