We describe an approach that uses causal and geometric reasoning to construct explanations for the purposes of the geometric features on the parts of a mechanical device. To identify the purpose of a feature, the device is simulated with and without the feature. The simulations are then translated into a “causal-process” representation, which allows qualitatively important differences to be identified. These differences reveal the behaviors caused and prevented by the feature and thus provide useful cues about the feature's purpose. A clear understanding of the feature's purpose, however, requires a detailed analysis of the causal connections between the caused and prevented behaviors. This presents a significant challenge because one has to understand how a behavior that normally takes place affects (or is affected by) another behavior that is normally absent. This article describes techniques for identifying such elusive relationships. These techniques employ a set of rules that can determine if one behavior enables or disables another that is spatially and temporally far away. They do so by geometrically examining the traces of the causal processes in the device's configuration space. Using the results of this analysis, our program can automatically generate text output describing how the feature performs its function.

We present a new representation that allows a rigid-body dynamic simulation to be described as a set of “causal-processes.” A causal-process is an interval of time during which both the behavior and the causes of the behavior remain qualitatively uniform. The representation consists of acyclic, directed graphs that are isomorphic to the flow of causality through the kinematic chain. Forces are the carriers of causality in this domain; thus they are central to the representation. We use this representation to compute the purposes of the geometric features on the parts of a device. To compute the purpose of a particular feature, we simulate the behavior of the device with and without the feature present. We then re-represent the two simulations as causal-processes and identify any causal-processes that exist in one simulation but not the other. Such processes are indicative of the feature's purpose. Because they are already causal descriptions of behavior, they can be directly translated into natural language descriptions of the feature's purpose. We have implemented our approach in a computer program called ExplainIT II.