Analytic and Learned Footstep Control for Robust Bipedal Walking

Abstract

Bipedal walking is a complex, balance-critical whole-body motion with inherently unstable inverted pendulum-like dynamics. Strong disturbances must be quickly responded to by altering the walking motion and placing the next step in the right place at the right time. Unfortunately, the high number of degrees of freedom of the humanoid body makes the fast computation of well-placed steps a particularly challenging task. Sensor noise, imprecise actuation, and latency in the sensomotoric feedback loop impose further challenges when controlling real hardware. <br /> This dissertation addresses these challenges and describes a method of generating a robust walking motion for bipedal robots. Fast modification of footstep placement and timing allows agile control of the walking velocity and the absorption of strong disturbances. In a divide and conquer manner, the concepts of motion and balance are solved separately from each other, and consolidated in a way that a low-dimensional balance controller controls the timing and the footstep locations of a high-dimensional motion generator. Central pattern generated oscillatory motion signals are used for the synthesis of an open-loop stable walk on flat ground, which lacks the ability to respond to disturbances due to the absence of feedback. The Central Pattern Generator exhibits a low-dimensional parameter set to influence the timing and the landing coordinates of the swing foot. <br /> For balance control, a simple inverted pendulum-based physical model is used to represent the principal dynamics of walking. The model is robust to disturbances in a way that it returns to an ideal trajectory from a wide range of initial conditions by employing a combination of Zero Moment Point control, step timing, and foot placement strategies. The simulation of the model and its controller output are computed efficiently in closed form, supporting high-frequency balance control at the cost of an insignificant computational load. Additionally, the sagittal step size produced by the controller can be trained online during walking with a novel, gradient descent-based machine learning method. While the analytic controller forms the core of reliable walking, the trained sagittal step size complements the analytic controller in order to improve the overall walking performance. The balanced whole-body walking motion arises by using the footstep coordinates and the step timing predicted by the low-dimensional model as control input for the Central Pattern Generator. Real robot experiments are presented as evidence for disturbance-resistant, omnidirectional gait control, with arguably the strongest push-recovery capabilities to date.