I've got the equipment, as I mentioned before, to record traces, but I'm spending a little more time making sure that I can do the inference on toy data. Right now I've got a small example demonstrating the ability to learn which activities are interrupted. On the left is what I have so far. I can do inference and learn with EM on this model.

Each activity has a node, which indicates the step in the activity which is currently underway. The data association variable shows which activity is currently making progress. The interruption variable fires when there is an interruption.

Some things that this model allows:
1. The ability to learn which activities get interrupted by which other activities.
2. The ability to learn which objects are expected to be observed in each step of an activity.

Some notable things which are missing from this model are:
1. The ability to preferentially return to a previously underway activity.
2. Relational smoothing over classes of activities.