I wonder if this modeling approach would be suitable for use with the Subsumption Architecture that Rodney Brooks developed around 1990. The ability to automatically identity features, adding new features as needed reminds me of the layered design process used in Subsumption. The idea of subsumption, in a nutshell, seems to be to create behaviors, starting with the most critical behavior first. As that behavior is perfected, another behavior is placed "on top". For example, have a layer that avoids objects and then a layer that navitages paths. One criticism of this approach is that there is no rigourous design process- the roboticist has to fumble through, largly unguided, to create his layers. Could this Markov Chain model, with it's ability to dynamically create new features be applied to Subsumption? Instead of giving the robot a 3d world to traverse and identify, give it samples of a successful task being performed (under control from the roboticist), for example, walking the hallway. Assuming the robot has already divided the physical space into Markov chains, as described in the paper, a simular algorithm could be used to divide the TASK into markov chains. In the paper, the robot learns to identify the location of hills in a sandbox. Could this algorithm be used to identify the completion of steps in a process?