Planning and execution of grasping motions on a humanoid robot

In this paper we present an approach for generating collision-free grasping motions and robustly execute them on a humanoid robot. The proposed MultiEEF-RRT algorithm for planning collision-free grasping trajectories exploits the enlarged goal space of a humanoid robot that results from the parallelized search of grasping trajectories for each arm. Here, multiple paths are searched simultaneously and the planner automatically chooses the first found solution. The reactive execution component operates on the planned C-Space trajectories and observes the movements in workspace with visual servoing approaches. The proposed algorithms do not rely on hand-eye calibrations, however it is possible to reliably execute given trajectories. The approach is fault-tolerant against changing execution speed, inaccurate sensor data and inexact executions of velocities. Since the hand and the target poses are visually tracked, the Cartesian error between the estimated position on a trajectory and the visually retrieved hand pose can be determined in workspace. This value is projected in the configuration space and used as a correction factor when calculating the joint velocities. We realized a grasping scenario with the humanoid robot ARMAR-III, where an object in front of the robot should be grasped. This demonstration shows how the proposed components play together to build a reactive and robust system integrating planning and execution of collision-free motions.