Howard Cannon

Modeling and identification of soil-tool interaction in automated excavation

In the process of automating an earthmoving machine, we have developed a model of soil-tool interaction that predicts resistive forces experienced at the tool during digging. The predicted forces can be used to model the closed loop behavior of a controller that servoes the joints of the excavator so as to fill the bucket. In this paper, we extend the state of the art in two ways. First, we present a reformulated version of the classical fundamental equation of earthmoving often used to model soil-tool interaction. The new model includes consideration of previously unaccounted phenomena in the interaction of an excavator bucket as it moves through soil. Secondly, given that soil properties can vary even within a work site, we present an online method to estimate soil parameters from measured force data. Finally, we show how the predicted resistive force is used to estimate bucket trajectories.

Multi-resolution planning for earthmoving

We suggest that planning for automated earthmoving operations such as digging a foundation or leveling a mound of soil, be treated at multiple levels. In a system that we have developed, a coarse-level planner is used to tessellate the volume to be excavated into smaller pieces that are sequenced in order to complete the task efficiently. Each of the smaller volumes is treated with a refined planner that selects digging actions based on constraint optimization over the space of prototypical digging actions. We discuss planners and the associated representations for two types of earthmoving machines: an excavator backhoe and a wheel loader. Experimental results from a full-scale automated excavator and simulated wheel loader are presented.

Models for Automated Earthmoving

We present a composite forward model of the mechanics of an excavator backhoe digging in soil. This model is used to predict the trajectories developed by a closed-loop force based control scheme given initial conditions, some of which can be controlled (excavator control parameters), some directly measured (shape of the terrain), and some estimated (soil properties). Since soil conditions can vary significantly, it is necessary that soil properties be estimated online. Our models are used to both estimate soil properties and predict contact forces between the excavator and the terrain. In a large set of experiments we have conducted, we find that these models are accurate to within approximately 20% and run about 10 times faster than real-time. In this paper we motivate the development of these models and discuss experimental data from our testbed.