Synthesis of Tactical Plans For Robotic Excavation

Abstract

This thesis describes an approach to synthesizing plans for robotic excavators. Excavation tasks range from loading a pile of soil to cutting a geometrically described volume of earth--for a trench or foundation footing. The excavation task can be stated in terms familiar to researchers in robotics and artificial intelligence: Transform the world from its current state to another state. Two important characteristics, however, distinguish the excavation domain. First, soil is deformable, and, hence, a complete state-space description of terrain to be excavated requires a high-dimensional representation. But, for the state-based representations that traditional planners use, very large spaces are simply infeasible. Second, the response of soil varies immensely, and in general, it is not possible to analytically describe the mechanics of soil motion and its interaction with tools. A soil's shear strength depends not only on its physical and chemical makeup, but also on factors such as the compaction experienced in the past. Thus a robotic excavator is forced to deal with approximate and incomplete models of its actions and their results. This research has developed and demonstrated a unified approach to deal with each of these issues. Instead of posing the planning problem in state space, this thesis employs a dual representation called "action space." An action space is spanned by parameters that abstract the actions that a robot excavator might perform. This formulation allows posing an excavation task as a problem of constrained optimization over the space of prototypical, one-step excavating plans. To reason about resistive forces encountered while digging, this work has developed a method that learns to predict resistive forces encountered during excavation. The action space representation allows incorporating goal configuration and geometric and force constraints within a single framework. Planning is thus reduced to finding a subset of plans that meet the constraints and optimize an appropriate performance measure. Over 400 experiments on a specially-developed testbed were conducted to validate the action-space/force-prediction approach. The testbed robot has a hydraulically manipulated shovel and sensors that measure both contact forces and terrain contour. Given a geometric specification, the current implementation can excavate a trench to predictable tolerances. Testbed results reinforce the notion that a force-based model is essential to successful robotic excavation. These findings also point to the necessity of a lower-level control mode that can modify excavation plans based on the nature of the terrain encountered.