Optimization and Learning-Based Planning and Control for Quadrupedal Manipulators

Although robotics has experienced huge advancement during the last 80 years, today’s robots still lack the intelligence to achieve human-like performance when physically interacting within unstructured environments. Quadrupedal mobile manipulators integrate the agility of legged locomotion with dexterous manipulation capabilities, enabling them to potentially navigate and interact autonomously in any environment that is accessible by humans. However, their deployment in real-world settings has been limited due to the extreme complexity of coordinating their motion during runtime in a computationally efficient, robust, and adaptive manner, particularly when generating contact-rich motion. In this thesis, we aim to tackle this by developing motion planning and control algorithms for quadrupedal manipulators. The first part of this thesis leverages the intuitiveness of optimization-based techniques. Firstly, we develop a Trajectory Optimization (TO) framework for motion planning in quadrupedal manipulators tasked with carrying heavy payloads on their arms. We adopt a payload-aware approach by incorporating the dynamics of both the robot and the payload in the optimization and opting for computational efficiency over model accuracy. This enables simultaneously planning Center-of-Mass (CoM) and payload manipulation motions, which reduces leg outstretching and allows the robot to maintain balance and maneuverability, especially during kinematically demanding movements like large steps. At the control level, a Whole-Body Controller (WBC) ensures that the TO plans are followed by generating whole-body joint trajectories. Our validation on hardware with payloads exceeding 15% of the robot’s mass while traversing non-flat terrain demonstrates the suitability of the approach for efficiently handling heavy payloads. At the next stage, a whole-body Model Predictive Control (MPC) strategy is developed to address the challenge of real-time whole-body planning and control in highly redundant quadrupedal manipulators. The MPC framework considers the full kinematics and a centroidal dynamics model to control a dual-arm quadruped with 37 actuated joints. Due to the whole-body information in the formulation, the MPC interfaces with the joint impedance controllers without the need for a WBC. The proposed MPC method is experimentally validated on a real robot, where it successfully solves complex tasks, such as picking up heavy objects from the ground while avoiding self-collisions and dynamically stepping with trotting and crawling gaits. These results underscore the importance of whole-body information in a predictive control approach. In the second part of the thesis, we explore a Reinforcement Learning (RL) approach for object goal pushing, where the quadrupedal manipulator must push an unknown object to a planar goal pose. We adopt a constrained RL approach where the control policy is trained in simulation and transferred to the hardware without any fine-tuning or other adjustments. The policy outputs joint space commands for the arm and 6D cartesian commands for the base. The latter are tracked by a pre-trained RL-based locomotion controller. While training, the proposed controller tries to maintain object balancing, which is important when the object has a small footprint or the flooring has high friction. The learned policy demonstrates robustness across objects with different masses, sizes, shapes, and materials despite only observing the object’s pose. Extensive testing on hardware reveals a success rate of at least 80 %, even in challenging scenarios, highlighting the controller’s effectiveness in contact-rich pushing. When on high-friction flooring, the controller adapts the pushing location to lower to prevent object toppling and to maintain effective manipulation.