The funnel algorithm and task level robot control

Abstract

Traditionally the translation of a given task into robot actuator commands involves three translation steps: (i) Task level planning; given the task find the manipulated object's required motion. (ii) Robot motion planning; given the manipulated object's required motion find the desired robot arm trajectory. (iii) Robot controller; given the desired robot arm trajectory find the robot actuator commands. Furthermore, the robot controller is fixed and internal and is typically not modified in accordance with the expected interaction between the robot arm and the objects it is manipulating. Thus the fixed robot controller is a major handicap which limits the possible robot interactions and performance during such interactions. In this paper we suggest that the objectives of the robot controller be "elevated" from attaining robot arm level objectives to attaining task level objectives. Consistent with this view, the following new three step Funnel algorithm is suggested for the translation of a given task into robot actuator commands: (i) Partition the task into a sequence of local sub-tasks and local regions. (ii) Synthesize a feedback control law to achieve each sub-task locally. (iii) Apply the local feedback control law to obtain the required actuator commands. It is expected that using this funnel approach both the robot system performance and functionality can be improved.