Department of Computer, Control and Management Engineering

In this paper, we consider the problem of generating appropriate motions for a torque- controlled humanoid robot that is assigned a multi-contact loco-manipulation task, i.e., a task that requires the robot to move within the environment by repeatedly establishing and breaking multiple, non-coplanar contacts. To this end, we present a complete multi-contact planning and control framework for multi-limbed robotic systems, such as humanoids. The planning layer works offline and consists of two sequential modules: first, a stance planner computes a sequence of feasible contact combinations; then, a whole-body planner finds the sequence of collision-free humanoid motions that realize them while respecting the physical limitations of the robot. For the challenging problem posed by the first stage, we propose a novel randomized approach that does not require the specification of pre-designed potential contacts or any kind of pre-computation. The control layer produces online torque commands that enable the humanoid to execute the planned motions while guaranteeing closed-loop balance. It relies on two modules, i.e., the stance switching and reactive balancing module; their combined action allows it to withstand possible execution inaccuracies, external disturbances, and modeling uncertainties. Numerical and experimental results obtained on COMAN+, a torque-controlled humanoid robot designed at Istituto Italiano di Tecnologia, validate our framework for loco-manipulation tasks of different complexity.