Comments on “A  Robotic Note”

Leland McCleary                               24 Oct.  2008

A Robotic Note is a useful outline of where we need to be heading.  As such, it will need to be elaborated upon until we have executable plans for each stage of the robot’s construction.  To stimulate discussion for this elaboration, I’m raising here some questions which the schematic architecture suggests to me in light of our prior discussions.  

     The design presented is: 

perception -> calculus 1 -> representation -> cognition -> calculus 2 -> motor function.

1)  My first question is:  what parts of this design must be “pre-fabbed” for the robot, and what parts must it build on its own?  And for the parts it must build on its own, what knowledge (or algorithm, or procedure) does it need to have installed so that it can build it?

     From what I’ve understood from the discussions, perception and motor function will have to be built in, both as hardware (for the material robot) and software (for both the material and the virtual robot).  In addition, I assume that calculus 1 and calculus 2 will also have to be provided to the robot so that it will know how to process perception, on the one hand, and know how to convert cognitive output into input for the motor function, on the other.  That leaves representation and cognition. 

     Much of the class-time discussion has revolved around how to get from a first-person representation to a third-person representation.  I wonder, in terms of the model above, whether that question can be paraphrased as “how to get from ‘raw perception’ to a third-person representation”.  If that’s the case, then I assume that calculus 1 should be powerful enough to produce a third-person representation with ‘raw perception’ as input.  Perhaps, then, calculus 1 would consist of multiple phases, the earlier of which would ‘pre-process’ perceptual input (perhaps into something like a first-person representation, if that concept continues to be relevant?) and the latter of which would convert that information into a full-blown third-person representation (assuming that such a conversion is possible, given the issues that have been raised concerning the role of interaction with other agents in the construction of third-person representations).

    From what I understand of what is said of intentional cognition, the project spaces containing an epistemic project schema and a narrative project schema will have to be pre-fabbed and installed.  That seems to be unavoidable.

2)  My second question (or rather, family of questions) is harder to articulate.  It has to do with the recursive and reflexive natures of cognition, action, perception and representation:  the fact that the flow of information in a cognitive robot will not be exclusively left-to-right (as in the diagram), which appears essentially “bottom-up”, but will also be looped back on itself, as learning and adaptation, such that intention affects perception, representations will adapt to favor certain cognitions over others, and actions will feed back into knowledge representations and intentional states.

     Some of this recursivity (but I think not all) is apparently built into the description of the narrative procedures, but not into the diagram, so perhaps it would be useful to try schematizing it.  Intentional cognition, as described, is a complex element, composed of project systems aimed at ‘knowing’ (epistemic projects) and ‘changing’ (narrative projects).  Here we have recursiveness described, at least in the narrative loop, with an ideal narrative program being written into a void representation and then translated into a realistic narrative program by projection into the remembered map (the representation) of the actual circumstances, resulting in an intentional representation that feeds into the motor calculus.  Presumably some kind of recursive loop would also be found in epistemic projects, assuming that “knowing” somehow involves changing the current state of the robot’s map of actual circumstances.  How these procedures are to be handled by the robot needs to be “fleshed out” (to use a most certainly inappropriate metaphor), and eventually transformed into programmable flow charts.

     One possible diagram of the narrative procedure might be something like the following:

Figure 1.

     The above graphic representation will have to be elaborated further to accommodate the description of how the linguistic narrative is achieved.  Here we have a recursive step-wise function built into the motor calculus (calculus 2), in which, as each step of the realistic narrative program is executed, successive perceptual scans of the robot’s world are performed, the map (representation 1) is updated, and the realistic narrative program is readjusted.  Although this was not explicitly stated, I assume that the recursion occurs after each step of the motor program’s execution, thus allowing for action on the part of the robot to change the context of situation.  For each loop for which there is external change and readjustment, a narrative episode is outputted in linguistic form.  It is not clear from the description what module of the program should output the episodes of the story, or, in fact, should convert them into linguistic accounts.  It is apparent from the diagram that calculus 2 must perform a complex set of functions which probably should be broken down into discrete modules.  Also, it is not apparent to me, if the robot is continuously scanning its environment, why it needs a scan specifically mandated by the motor calculus for the purposes of detecting narrative situations, or whether its “continuous scan” is always necessarily part of a narrative loop.

     Figure 2 attempts to diagram these relations (with the project space of Figure 1 simplified).

Figure 2.

3)  My third question has to do with how the internal language of the robot, in which the robot constructs its narratives and its knowledge out of the mismatches between its current states and current intentions, becomes transparently comprehensible to its human interactants.  I suspect this will only be possible if a language interpretable by humans is given to the robot by its human builders, in a kind of translation module.  As we have discussed elsewhere, I fail to see how this solves the symbol grounding problem.