Whole-Body Planning and Control for Multi-Contact Loco-Manipulation

Abstract

As we advance into an era increasingly characterized by the integration of autonomous systems into our everyday lives, legged mobile manipulators stand at the forefront of this shift, embodying the critical challenge of enhancing the motor skills of physical robots to become useful assistants in various aspects of human life. This dissertation focuses on advancing the planning intelligence and control capabilities of these systems, equipping them with the versatility and robustness to solve complex real-world loco- manipulation tasks involving multiple contact interactions. Initially, the research delves into Contact-Implicit Optimization (CIO), introducing a multiple shooting scheme and leveraging a fast structure- exploiting solver to efficiently generate multi-contact plans for a robotic arm performing non-prehensile manipulation tasks. Recognizing the computational challenges associated with CIO when considering high-dimensional systems, the thesis transitions towards an Optimal Control (OC) framework for switched-systems to unify the treatment of dynamic locomotion and manipulation tasks. We demonstrate that by opting for a sufficiently-rich reduced-order model that encapsulates essential dynamic couplings, we achieve a real-time generation of whole-body trajectories for a quadrupedal mobile manipulator. We also elaborate on the underlying Nonlinear Model Predictive Control (NMPC) scheme that was used, and show how it was further enhanced with an Augmented-Lagrangian (AL) approach to enforce various path constraints. In pursuit of automating the generation of holistic loco-manipulation behaviors with minimal manual guidance, the research introduces a new perspective that is centered around modeling multi-contact loco-manipulation tasks as integrated Task and Motion Planning (TAMP) problems. As a result, a computationally tractable bilevel optimization problem emerges, efficiently solved by a sampling-based bilevel search algorithm that combines the strengths of different planning techniques. This thesis extends its contributions by focusing on the design of loco-manipulation policies capable of executing these behaviors robustly, withstanding significant modeling mismatches and external disturbances. This effort is based on a task-agnostic unified Markov Decision Process (MDP) formulation, and leverages demonstration-guided Deep Reinforcement Learning (DRL) to achieve this goal. A series of hardware experiments are conducted that validate the dynamic feasibility and robustness of our proposed methodologies, showcasing their practical applicability across a range of complex loco-manipulation tasks. Through these contributions, the thesis not only advances the state-of-the-art but also lays a foundational framework for future research in multi-contact planning and control for legged loco-manipulation.