Addressing Flexibility in Biped Locomotion with Robust Control and Closed-loop Model-Predictive Control

While the last two years have seen the rise of many quadruped robots with excellent locomotion capabilities, biped robots are still limited, as they are evolving in a stability zone of reduced size. When transferring a locomotion controller from simulation to reality, modeling errors are then difficult to compensate with feedback only. This consequently imposes drastic constraints on the hardware design. In this paper, we propose to consider the simulation-to-reality gap by designing a robust locomotion controller. The robustness is obtained by a quantitative analysis of uncertainties, leading to bounds on its effects. As these bounds are compatible with the robot constraints, we propose a robust controller able to produce dynamic walking gaits. Feedback is obtained through the robust controller, acting as a balance stabilizer, and through a closed-loop model-predictive controller modeling the centroidal dynamics. We apply the proposed scheme to control the locomotion of the humanoid robot Talos, whose hip is mechanically flexible by design. We demonstrate in simulation the importance of the robustness to handle this situation and show its application in various scenarios in stairs and subject to important disturbances.