Cognitive Robotics

 

Wish list

Page history last edited by Leland McCleary 10 mos ago

 

Wish list of skills

for a symbolically-grounded cognitive robot

 

Leland McCleary

[Ver. 0.02 – 06 Oct 2008]

 

This list is a collection of cognitive skills that have been assumed or attested in humans, in this first version based largely on Tomasello (1999). Some of them have been tested for in non-human primates, and some have been set as goals in robot research, with varying results.  It is an open list which may be expanded as further skills are identified.  At the moment the list goes no further than eye gaze and pointing (as incipient symbolization) and does not include the use of other gesture or vocalization.

 

The list is very roughly in order of ontological appearance, though it is impossible to chronologize  the list for a number of reasons.  First, the categorizations of skills originate in diverse types of research which produce results that are not strictly comparable.  Further, the items are at varying levels of granularity, which means that many of the items may in fact be sub-items of others.  Finally, while certain relations of priority or enablement may appear to hold among some of the items, their appearance may nevertheless appear to occur simultaneously.

 

An example of how the ordering of the items both does (to some extent) and does not hold is the item “that Self is distinct from Other”.  Gold & Scassellati (2007) have shown that a robot can learn to distinguish Self from animate Other (and from objects) based on internal motor state representations.  While this robot skill does not require an item earlier on the list, “that Self is an intentional being”, it also does not require “that Self can jointly engage with Others”, which supports the latter’s appearing later on the list.

 

Another example is the skill “that Self and Other can jointly attend”, investigated for robots by Scassellati (1996).  Scassellati, expanding on Baron-Cohen (1995) , concludes that being able to jointly attend can be achieved in four steps, among which are gaze monitoring (“that Self can attend to face, eyes and gaze of Other”), interpolation of gaze (“that Self can discover object of Other’s attention by ‘following’ gaze”) and pointing (“that pointing … can influence knowledge of other by directing attention to some object”).  In the list below, the first two skills precede the target behavior and the third follows, an ordering subject to further refinement of the categories.

 

What cognitively modern children must know (or learn)

·         …that seeing is knowing

·         …that objects exist in the world

·         …that objects can move (and remain the same objects)

·         …that Self can move and move its parts in controlled ways

·         …that Self can influence objects by moving its parts (can “make things happen”)

·         …that objects can influence (effect) other objects

·         …that Self can jointly engage with objects

·         …that Self can control shift in (choose) focus of attention

·         …that Self can achieve its ‘goals’ by moving its parts

·         …that Self is an intentional being (that its goals are separate from its behavioral means; innate, shared with other primates)

·         …that Self can focus on final goals and mediating goals concurrently

·         …that Self can influence objects as mediating goals to achieve a final goal

·         …that objects are different from Others

·         …that Others are animate beings with powers of self-movement that behave in certain ways, that make things happen

·         …that Self is distinct from Other

·         …that Self can attend to face, eyes and gaze of Other

·         …that Self can jointly engage with Others (dyadic engagement; jointly gazing at each other and interacting)

·         ...that Self can imitate action of Other (dyadic mimicry)

·         …that Others are ‘like me’ (identification: uniquely human, innate capacity?)

·         …that Others are intentional beings (with goals, attention and behavioral  strategy)

·         …that Others’ seeing is Others’ knowing

·         ...that Self can discover object of Other’s attention by ‘following’ gaze

·         …that Self and Other can jointly attend

·         …that Self can influence Other as mediating goal to achieve a final goal

·         …that Self can engage with Other mediated by object (triadic engagement; joint engagement)

·         …that Self can imitate intentional action of Other (imitative learning)

·         …that Self can benefit from intentional affordances of artifacts and attentional strategies of Others

·         …that Self can influence object of Other’s attention by gazing at it (and monitoring Other’s gaze)

·         …that pointing can influence Other by directing attention to object of Self’s goal (Imperative pointing)

·         …that pointing (or showing) can influence knowledge of other by directing attention to some object (Declarative pointing and showing)

 

References

BARON-COHEN, S.  (1995). Mindblindness. MIT Press.

GOLD, Kevin; SCASSELLATI, Brian. (2007). A Bayesian robot that distinguishes “Self” from “Other”. Proceedings of the 29th Annual Meeting of the Cognitive Science Society (CogSci2007). Nashville, Tennessee. Available at: <http://cs-www.cs.yale.edu/homes/scaz/papers/Gold-CogSci-07.pdf>.

SCASSELLATI, Brian. (1996). Mechanisms of shared attention for a humanoid robot. AAAI Fall Symposium “Embodied Intelligence.” Cambridge, MA. Oct. 1996. Available at: <http://cs-www.cs.yale.edu/homes/scaz/papers/scaz-autism.pdf>.

TOMASELLO, Michael. (1999). The cultural origins of human cognition. Harvard U. Press.

 

Comments (5)

profile picture

Matthew McCroskey said

at 2:52 pm on Oct 10, 2008

This list is a great start; perhaps we could use this list (or if we wish, use an expanded form of it that we devise) to pick what capabilities we feel are most vital to have in our robot. I understand your disclaimer that these capabilities obviously do not represent everything that a cognitively modern child (and ideally, a cognitively modern robot) must know, and that they do not appear in strictly chronological or hierarchical order, but I still feel that we could use this list or something like it to see what capabilities we need to enable the capabilities we desire.

profile picture

Kevin P. Barry said

at 12:16 am on Dec 14, 2008

This seems to be a very comprehensive list, and it takes into account the majority of having a grounded view of the world for the purposes of having a purpose. Unfortunately, not all humans readily utilize all of these characteristics; in fact, I doubt that half of humans actually "fully appreciate" these abilities for positive ends. In that vain, we might define our requirements in terms of "strength of personality," or balance of reality-grounded traits given in this article. Given that a human can indeed guide one's own attention and intention, these traits cannot remain in consistent proportion over time because all won't be important at all times.

Rather than a serialized list of possibilities, the list has great potential for representation as a recursive set of principles. For example:

Sentients ⊊ Things
Objects ≣ Things - Sentients
Self ∈ Sentients
Sentients "can move" Movable Objects | Movable Objects ⊂ Objects
Sentients "have" Attentions
Sentients "can influence" Attentions
∃(Attentions "can be influenced") ∀ Sentients
Nearby Sentients "can be influenced" | Nearby Sentients ⊊ Sentients
Self ∈ Nearby Sentients
Self ∉ Sentients - Self ≣ Others
Sentients "have" Knowledge
Seeing ⊊ Knowledge
Seeing ⊊ Attentions (∴ only some "seeing" can be influenced)
... etc.

This is just a start and it doesn't include everything. There is probably a better model to represent it with, but this is a start. I'll see how else I can represent it or compile it to reduce possible duplications or inconsistencies. In other words, extract the fundamentals that make all of these true.

Kevin Barry

profile picture

Kevin P. Barry said

at 12:30 am on Dec 14, 2008

Something you might have noticed in the abstracted example model in my other post is that it allows for the possibility of some "seeing" to *not* be "knowledge," which is a key principle of attention. Something I wish I would have incorporated was the fact that "self" is a member of "nearby sentients" of all other "nearby sentients," which would have tied in a whole lot of other things. I'm still not sure how that can gracefully be represented.

Kevin Barry

profile picture

Leland McCleary said

at 9:46 pm on Jan 31, 2009

Kevin, I think the kind of formalization you're proposing is exactly the right move toward capturing the non-linear and embedded interrelatedness of the skills. Still, it might be misleading to represent the list as a "grammar" of relationships that capture all and only the attested skills, without representing somehow a developmental (temporally evolving) dimension. I'm sure this is tricky, since with each accretion the environment within which previously acquired skills operate changes, making possible new relationships among old elements. I'm really just groping here, so I'm afraid I can't be more specific.

I'd like to understand better your point about "'strength of personality', or balance of reality-grounded traits". Intuitively, I'd say that part of the reason some of these traits seem to lose their (grounded) bearing for most humans (I think "half" is being generous) is that the further into the semiotic realm we are driven by these mechanisms, the more tenuous grounding becomes; and of course it's just that "forgetting" that is essential to the trick of making our social semiotic world seem real.

Which reminds me that somewhere in an expanded list of cognitive skills will have to begin to appear the skills allowing the forgetting.

Just groping.
Leland McCleary

profile picture

Kevin P. Barry said

at 8:12 pm on Feb 2, 2009

I do think temporal evolution needs to be a consideration. Hypothetically speaking, it could be that each new rule of this sort of grammar is learned as it applies to it's association with something specifically-known, and its global applicability is reaffirmed as it applies to new situations. Not that they're perceived as universal initially; it could be that universality is retroactively applied as it proves plausible. I don't suspect it too controversial to proclaim that the universality precedes its knowledge, and that one could wait for a child to make each realization a step at a time. Upon acceptance of these realizations, one might say the person has reached "full" sentience, but until that time they don't necessarily realize their ability to influence and to be influenced. I don't know that this process (begging the question) is complete until a few years into adulthood, but I don't think that any benefit arises from a machine learning these ideas in real time. These ideas come to be self-evident in the minds of humans, yet they arise out of utter chaos exant in the early neural system. Because we fail to control the free wills of actual humans, we must concede that some level of predictability must exist in intelligent machines; otherwise, they'd be uncontrollable.

Kevin Barry

You don't have permission to comment on this page.