Dynamic interaction of robotic systems, humans and animals with their environment through the creation of intermittent physical contacts is at the heart of any locomotion or manipulation behavior. Yet, the generation of robust behaviors with intermittent contacts in a dynamically changing unknown environment remains a fundamental challenge for robotics research. We argue that our current inability to plan and control behaviors with contacts that are robust to external disturbances and environmental uncertainty constitute a major inhibitor to the deployment of robotics in unknown environments.
In this line of research, we investigates algorithms that explicitly take into account uncertainty stemming from unmodeled dynamics or measurement noise. In particular, we design numerically efficient algorithms for output feedback risk-sensitive optimal control and stochastic model-predictive control. We apply these ideas to the computation of optimal movements and impedance profiles that are robust to physical interactions.