Robotic Mobile Manipulation via Adaptive and Learning-Based Model Predictive Control

Abstract

The use of dynamically stable mobile manipulators is expanding beyond controlled research laboratories and into the real world. However, autonomous manipulation skills are still specialized to individual tasks and can handle limited variations in the objects’ physical properties, which prevents robot deployment in unstructured human environments. This thesis focuses on whole-body motion planning and control for dynamically stable mobile manipulators and on providing controllers with real-time adaptation to changes in the robot dynamics due to the interaction with objects. Dynamically stable mobile manipulators, i.e., actively-balancing mobile robots equipped with robotic arms, are potentially very useful for working in environments designated for humans. However, their agility and compliance come at the cost of a high control complexity. Traditional control strategies treat the locomotion and manipulation problems separately, requiring additional heuristics to achieve whole-body coordination. Moreover, inverse-dynamics-based controllers do not consider the system’s future evolution, which is essential for balance control. On the other hand, in this thesis we propose a whole-body motion planning and control formulation based on Model Predictive Control (MPC). Our method leverages the full robot dynamics and jointly optimizes for balancing, base tracking, end-effector tracking, and environment interaction. We validate the proposed whole-body MPC controller in extensive experiments with a ball-balancing manipulator. When the robot dynamics is not known precisely or when manipulating new objects, model uncertainties can severely undermine the performance and generality of MPC. To tackle this problem, we propose two online adaptation schemes for the objects’ parameters used in the system dynamics of MPC, which we showcase in a door-opening and an object-lifting task with a ball-balancing manipulator. Although we initially model the external environment as a linear system, a more descriptive representation is necessary for more complex manipulation tasks or in the presence of uncertainties in the robot dynamics. Thus, we propose to approximate the model error as a linear combination of trigonometric basis functions. Assuming that the underlying structure of the dynamics does not vary significantly when the robot performs similar manipulation tasks, we learn the hyperparameters of the basis functions from data collected during related experiments, e.g., by letting the robot open doors with different stiffness coefficients. When executing a new task, the hyperparameters of the basis functions are kept fixed while the linear parameters are adapted online. We test the resulting multi-task learning MPC controller in simulations and hardware experiments, and present extensive comparisons against other adaptive MPC controllers. Finally, to obtain better tracking performance despite parametric uncertain- ties, we incorporate the Control Lyapunov Function (CLF) constraint derived in adaptive control for robot manipulators to the set of inequalities of the optimal control problem. Thus, we obtain an adaptive controller that combines the advantages of CLFs and MPC, yielding an improved performance during the robot’s interaction with unknown objects and a reduced dependence on the tuning of the MPC prediction horizon. We demonstrate the advantages of the proposed method with respect to several baselines, and we validate it in hardware tests on a quadrupedal robot carrying bricks and pulling heavy boxes.