A Whole-Body Stack-of-Tasks compliant control for the Humanoid Robot COMAN Alessio Rocchi * Enrico Mingo Hoffman * Edoardo Farnioli * † Nikos G. Tsagarakis * Abstract—A fundamental aspect of controlling humanoid robots is the capability to use the entire body to perform tasks. In this paper we present an ongoing work to add this capability to the compliant humanoid robot COMAN, designed at the Italian Institute of Technology. Our control architecture is composed by a high level, whole-body inverse kinematic solver and a decentralized, low level, joint impedance control. Such architecture allows to regulate impedance using different strategies maintaining a high level of robustness and it has been developed to perform rescue operations in disaster scenarios. Keywords—Whole-Body Control, Humanoid Bipedal Robot, Stack of Tasks, Compliance I NTRODUCTION The recent DARPA Robotics Challenge showed all the difficulties that arise when advanced research topics need to be connected to practical tasks. In particular these tasks have to be performed in unstructured environments where classical stiff robots and controllers are not well suited. The main problem in real world scenarios are the uncertainties associated with the robot state and the environment state. These uncertainties make the interaction between the robot and the environment potentially dangerous. On the other hand, stiff controllers such as the classical PID position control are robust and can be implemented in a decentralized way. We think decentralized control schemes are fundamental in robots that have to operate continuously, without the possibility to be turned off. Let’s take for example a semi-autonomous and tele-operated humanoid bipedal robot: it has always to keep balance also in the case in which a recoverable hardware or software failure occurs on its on-board computer, or the communication with the robot is temporarily severed. Such scenario would require a restart of the control modules: if the robot is in a stable configuration the decentralized joint impedance control scheme can maintain the standing posture in the robot up to the end of the restart procedure, where a centralized control scheme could not be restarted without first bringing the robot to a safe rest configuration. With these considerations in mind, pure torque controllers such as in the case of operational space control schemes [8] are not well suited to systems where robustness is critical, but at the same time stiff position controllers are not suited for the interaction as well. A trade-off between the two approaches can be represented by a combination of a centralized pure kine- matic control together with joint impedance control. In fact, the latter can be implemented in a decentralized way and it adds compliance to the system allowing the possibility to regulate the impedance both at the joint level and at the Cartesian level through conservative congruence transformation [9]. Fig. 1. COMAN is a humanoid bipedal robot equipped with series elastic actuators (SEA) and torque controlled. It has 29 DOFs, 4 Force/torque sensors (two in the ankles and two in the arms) and one IMU placed in the waist. In this paper we present the control architecture that we are developing based on a Cartesian whole-body trajectory generator and a joint impedance controller for the humanoid bipedal robot COMAN [13] (Figure 1), developed at the Italian Institute of Technology. First we introduce some of the work done in the Whole-Body Inverse Kinematics/Dynamics compliant control then we introduce our methodology, finally we show some experimental results obtained on the humanoid robot COMAN. RELATED WORKS Kinematic and dynamic inversion are well known problems in robotics. In general, given some tasks specified in Cartesian position, velocity, accelerations or forces, we want to find the joints position, velocity, accelerations or torque that realize those tasks. Many solutions to this problem have been pre- sented for single and multiple kinematic chains [12] [14]. An interesting subset of these algorithms are the ones based on numerical optimization. If on one side numerical optimization could be less efficient than other algorithms at solving the kinematic inversion problem, on the other side they allow for the explicit introduction of unilateral/bilateral constraints in the inverse kinematic/dynamic problem, which are fundamental when working on the real hardware.