Offline and Online Planning and Control Strategies for the Multi-Contact and Biped Locomotion of Humanoid Robots

In the past decades, the Research on humanoid robots made progress forward accomplishing exceptionally dynamic and agile motions. Starting from the DARPA Robotic Challenge in 2015, humanoid platforms have been successfully employed to perform more and more challenging tasks with the eventual aim of assisting or replacing humans in hazardous and stressful working situations. However, the deployment of these complex machines in realistic domestic and working environments still represents a high-level challenge for robotics. Such environments are characterized by unstructured and cluttered settings with continuously varying conditions due to the dynamic presence of humans and other mobile entities, which cannot only compromise the operation of the robotic system but can also pose severe risks both to the people and the robot itself due to unexpected interactions and impacts. The ability to react to these unexpected interactions is therefore a paramount requirement for enabling the robot to adapt its behavior to the task needs and the characteristics of the environment. Further, the capability to move in a complex and varying environment is an essential skill for a humanoid robot for the execution of any task. Indeed, human instructions may often require the robot to move and reach a desired location, e.g., for bringing an object or for inspecting a specific place of an infrastructure. In this context, a flexible and autonomous walking behavior is an essential skill, study of which represents one of the main topics of this Thesis, considering disturbances and unfeasibilities coming both from the environment and dynamic obstacles that populate realistic scenarios.  Locomotion planning strategies are still an open theme in the humanoids and legged robots research and can be classified in sample-based and optimization-based planning algorithms. The first, explore the configuration space, finding a feasible path between the start and goal robot’s configuration with different logic depending on the algorithm. They suffer of a high computational cost that often makes difficult, if not impossible, their online implementations but, compared to their counterparts, they do not need any environment or robot simplification to find a solution and they are probabilistic complete, meaning that a feasible solution can be certainly found if at least one exists. The goal of this thesis is to merge the two algorithms in a coupled offline-online planning framework to generate an offline global trajectory with a sample-based approach to cope with any kind of cluttered and complex environment, and online locally refine it during the execution, using a faster optimization-based algorithm that more suits an online implementation. The offline planner performances are improved by planning in the robot contact space instead of the whole-body robot configuration space, requiring an algorithm that maps the two state spaces.   The framework proposes a methodology to generate whole-body trajectories for the motion of humanoid and legged robots in realistic and dynamically changing environments.  This thesis focuses on the design and test of each component of this planning framework, whose validation is carried out on the real robotic platforms CENTAURO and COMAN+ in various loco-manipulation tasks scenarios.