Advancing Loco-Manipulation and Interaction of Quadruped Manipulators: Integrating Control, Perception, and Planning

As the integration of robotic systems into our daily lives intensifies, there is a growing interest in legged manipulators. Their appeal largely lies in their ability to merge the capabilities of stationary robotic arms and mobile units with legs into a cohesive system. Optimization-based techniques have demonstrated the potential for exploring the duality between locomotion and manipulation, generating coordinated, dynamically feasible whole-body motions. However, unimodal planning is only a step towards autonomy. To advance from isolated tasks to complex, real-world operations, legged manipulators must integrate control, perception skills, and long-term planning. This thesis presents control solutions across these three levels, aimed at planning and executing sequences of motions, with reactive and compliant behavior based on visual feedback and force control. The design of these algorithms has the goal of pushing what the current state of the art is capable of doing in terms of dynamic-legged loco-manipulation. In Chapter 2, we focus on control and we present a Whole-Body Control that integrates a Cartesian impedance controller to coordinate tracking performance and desired compliance for the robot base and manipulator arm. The controller is formulated through an optimization problem using Quadratic Programming to impose a desired behavior for the system while satisfying friction cone, unilateral force constraints, joint and torque limits. The presented strategy decouples the arm and the base of the platform, enforcing the behavior of a linear double-mass spring damper system, and allows to independently tune their inertia, stiffness and damping properties. We perform an analysis of the controller capabilities in rendering impedances, considering inertia shaping for the base as well as for the arm end-effector. We investigate how a trotting gait affects the accuracy in rendering the desired impedance and how the gait parameters also affect the impedance tracking. Chapter 3 focuses on integrating perception with control in a reactive way. The thesis proposes a kinematically decoupled control approach that integrates an \gls{ibvs} scheme and impedance control. The control approach maps the visual task only on the wrist, exploiting low-inertia links for fast motion and reactiveness, and the rest of the kinematic chain for less demanding arm positioning. A heuristic sequence of behavior and control signals for the Search, Approach and Grasp of an object placed at an unknown location is proposed. Across the different phases, the visual feedback error is mapped to base and arm commands to approach and grasp the object. Chapter 4 focuses on integrating control with task and motion planning. We adopt a path optimization perspective of the Task and Motion Planning problem, using the Logic-Geometric Programming formalism to solve for the discrete sequence of actions and their corresponding motion plans. We search over actions represented as nodes in an expanding tree, and we run multiple calls of K-order Markov Optimization to test logically feasible sequences. The actions come defined with logical predicates and a set of costs/constraints interpreted by K-order Markov Optimization. K-order Markov Optimization optimizes in a receding horizon fashion body configurations considering the robot kinematics. Additionally, the motion planning is performed over a simplified template for a legged manipulator, comprising of a 6 multi-degree of freedoms floating base and a robotic manipulator. The re-planning gives robustness properties to the robot motion control to deal with dynamic environments. Finally, body configurations that can lead the robot to the final state/scene are time-parametrized and executed with the Whole-Body Control, to guarantee the dynamic feasibility of the kinematic plans. The presented strategies are validated on simulation and on hardware, using the quadruped robots HyQ (90 kg) and HyQReal (140 kg) both equipped with a Kinova Gen3 robotic arm. The thesis shows that with the help of these strategies, legged manipulator robots are one step closer to being fully autonomous, more perception-aware, and more robust for real-world applications.