Estimation-based control for humanoid robots

As sensor, actuator and processor technology continues to improve, humanoid robots have become more common in both academic and industrial environments as general-purpose platforms capable of performing a wide variety of automated tasks. These robots have the potential to operate in complex environments built for humans; their form factor allows them to navigate buildings, utilize existing tools and perform cooperative tasks with humans, making them ideal for disaster recovery and household labor among many other applications. However, the challenge of operating autonomously in unknown environments involves obtaining accurate estimates of the robot’s state by fusing information from on-board sensors and using these estimates for control in ways which allow robustness to uncertainty and disturbances. In this work, we propose methods for estimating important states of humanoid robots and evaluate the role of sensory information and state estimation in executing behaviors on a torque-controlled humanoid. First, we discuss estimation of the floating base pose of the robot in the world frame using only onboard sensing and kinematics models. This estimator takes into account contact switching and is shown to exhibit desirable observability characteristics through a nonlinear observability analysis. While the base pose and its velocity are crucial in computing joint torques via inverse dynamics, the configuration-dependent center of mass (COM) and its dynamics (the momentum of the system) is more-often the quantity of interest for planning and control. However, robot dynamic models are never perfect; we thus introduce estimators using the momentum dynamics of the robot with measured contact wrenches to estimate configuration-dependent center of mass and momentum offsets as well as external wrenches applied about the COM. We propose to evaluate the utility of this estimator for online replanning in a momentum model predictive control (MPC) framework, thereby indirectly using measured contact wrenches for feedback. We next introduce methods for computing joint velocities and accelerations from link-mounted IMUs and knowledge of kinematics, avoiding numerical differentiation of noisy joint angle sensors. This information is fused in several estimators and used to increase the maximum stable joint feedback gains; we propose to utilize inertial sensor-based estimates of joint derivatives in our whole-body control framework to achieve better Cartesian damping control and thus improve tracking and robustness in the system. Finally, we discuss the importance of contact estimation and the role it plays in whole-body control, introducing an unsupervised learning method for estimating contact probability from proprioceptive sensor data. This continuous measure of contact is shown to improve kinematics-based base pose estimation and therefore improve closed-loop control when walking on simulated rough terrain. These estimation methods are discussed and evaluated in the context of optimization-based inverse dynamics control for locomotion.