Motion planning and bimanual coordination in humanoid robots 1 Pietro MORASSO a,b,2 , Vishwanathan MOHAN a , Giorgio METTA a,b , Giulio SANDINI a,b a Italian Institute of Technology, Genoa, Italy b Department of Informatics, Systems, Telematics, University of Genoa, Italy Abstract. Humanoid robots have a large number of “extra” joints, organized in a humanlike fashion with several kinematic chains. In this chapter we describe a method of motion planning that is based on an artificial potential field approach (Passive Motion Paradigm) combined with terminal-attractor dynamics. No matrix inversion is necessary and the computational mechanism does not crash near kinematic singularities or when the robot is asked to achieve a final pose that is outside its intrinsic workspace: what happens, in this case, is the gentle degradation of performance that characterizes humans in the same situations. Moreover, the remaining error at equilibrium is a valuable information for triggering a reasoning process and the search of an alternative plan. The terminal attractor dynamics implicitly endows the generated trajectory with human-like smoothness and this computational framework is characterized by a feature that is crucial for complex motion patterns in humanoid robots, such as bimanual coordination or interference avoidance: precise control of the reaching time. Keywords: Motion planning, Humanoid robot, Attractor dynamics, Terminal attractors. Introduction Humanoid robots have a large number of “extra” joints, organized in a humanlike fashion with several kinematic chains. Consider, for example, Cog [1] with 22 DoFs (Degrees of Freedom), DB [2] with 30 DoFs, Asimo [3] with 34 DoFs, H7 [4] with 30 DoFs, iCub [5] with 53 DoFs. When a finger touches a target, the elbow might be up or down and the trunk may be bent forward, backward or sideways. Thus an infinite number of solutions are available to the motor planner/controller. This redundancy is advantageous because it enables a robot to avoid obstacles and joint limits and attain more desirable postures, for example when it is not sufficient to simply tap a target because a precise force vector must be applied to the touched object. From a control and learning point of view, however, redundancy also makes it quite complicated to 1 IOS press, series KBIES (Knowledge-Based Intelligent Engineering Systems); subseries of "Frontiers in Artificial Intelligence and Applications"; book title "Computational Intelligence and Bioengineering"; editors: F. Masulli, A. Micheli, A. Sperduti (2009). 2 Corresponding author: Pietro Morasso, University of Genoa, DIST, Via Opera Pia 13, 16145 Genoa, Italy. Email: pietro.morasso@unige.it