Navigation Among Movable Obstacles

Abstract

Robots would be much more useful if they could move obstacles out of the way. Traditional motion planning searches for collision free paths from a start to a goal. However, real world search and rescue, construction, home and nursing home domains contain debris, materials clutter, doors and ob jects that need to be moved by the robot. Theoretically, one can represent all possible interactions between the robot and movable ob jects as a huge search. We present methods that simplify the problem and make Navigation Among Movable Obstacles (NAMO) a practical challenge that can be addressed with current computers. This thesis gives a full development cycle from motion planning to implementation on a humanoid robot. First, we devise a state space decomposition strategy that reasons about free space connectivity to select ob jects and identify helpful displacements. Second, we present controls for balance and manipulation that allow the robot to move ob jects with previously unknown dynamics. Finally, we combine these results in a complete system that recognizes environment ob jects and executes Navigation Among Movable Obstacles. Our continued work in NAMO planning has focused on three topics: reasoning about ob ject interaction, three dimensional manipulation and interaction with constrained ob- jects. This thesis presents the computational and theoretical challenges that arise from these elaborations of the NAMO domain. In each case we introduce extensions to our algorithms that respond to the challenge and evaluate their performance in simulation. All the methods presented in this thesis not only solve previously unsolved problems but also operate efficiently, giving real-time results that can be used during online operation.