Dottorato in AUTOMATICA, BIOINGEGNERIA E RICERCA OPERATIVA -
ABRO

Titolo della tesi: Motion Planning Techniques for Humanoid Robots

Thanks to their human-like structure, humanoid robots have the potential for accomplishing complex tasks requiring legged locomotion and/or dual-arm manipulation in both structured and unstructured environments. To successfully fulfil such tasks, appropriate humanoid motions must be generated. Planning these motions is particularly challenging for humanoid robots because of their peculiar characteristics. First, their high number of degrees of freedom makes planning computationally expensive. Second, they can displace their base only through stepping or acyclic multi-contact motions. Third, they must maintain balance at all times. This thesis addresses the motion planning problem for humanoid robots in various contexts. We start by considering the general problem of planning whole-body motions for a humanoid robot that must execute a task implicitly requiring locomotion in an environment populated by static obstacles. For this problem we propose a complete framework that can incorporate tasks of different nature (i.e., navigation, reaching, manipulation and visual tasks) for planning both in case of known and unknown environments. One of the advantages of humanoids is the possibility of moving through complex environments by stepping over or onto obstacles. To this end, we propose an integrated method for planning and executing humanoid motions on uneven ground. It is composed by two modules: an offline footstep planner and an online gait generator. For the first module we propose two possible randomized strategies that can efficiently compute feasible and optimal footstep plans, respectively. In many practical applications, it might be allowed to abandon the task in favor of collision avoidance. For cases in which the robot is assigned a soft task of this type, we present an opportunistic strategy for planning motions that, differently from other approaches, allow the robot to perform the assigned task for as long as possible, and deviate from it only when strictly needed to avoid a collision. The method is first discussed with regard to a generic free-flying robot, and later extended to the case of humanoid robots. More complex tasks that a humanoid robot can potentially fulfil require to sequentially establish with the environment multiple contacts involving not only the feet as in basic biped locomotion. For this problem we propose a multi-contact motion planner that thanks to its randomized nature avoids any kind of precomputation or heuristics design that are usually required with existing search-based techniques. Finally, we consider the problem of safe coexistence between human and humanoids. In this context, reactive planning capabilities are essential. We describe a complete framework for the safe deployment of humanoid robots in environments containing humans, where several safety behaviors are activated and deactivated through a state machine according to information coming from the robot sensors.