Improving Robustness in Complex Tasks for a Supervisor Operated Humanoid

Complex manipulation tasks in uncontrolled environments are challenged with errors from multiple sources that can prevent successful completion. We describe a framework for the decomposition of a complex task into behaviors for a supervisor controlled robot. A classification of behaviors based on the actors and dominant motion is used to analyze the success rate. Three methods to improve robustness are presented: reduction in length of robot-environment manipulation by using robot-only prepositioning behaviors, behavior definition using environmental constraints, and supervisor fine tuning during sub-task switching. We show the application of this framework for the wall task in the DARPA Robotics Challenge. The framework produces a robust successful implementation of the wall task with a duration of less than 10 min.