Moderated by Stephen Muggleton.

Nils J. Nilsson

Teleo-Reactive Programs and the Triple-Tower Architecture

The article mentioned above has been submitted to the Electronic Transactions on Artificial Intelligence, and the present page contains the review discussion. Click here for more explanations and for the webpage of theauthor, Nils J. Nilsson.

Overview of interactions

N:o Question Answer(s) Continued discussion
1 15.10  Anonymous Referee 1
2 15.10  Anonymous Referee 2

Q1. Anonymous Referee 1 (15.10):

Recommendation: The paper can be accepted for ETAI after minor revisions. No further refereeing round is needed.

This paper describes an agent architecture that permits a combination of reactive behaviour and planning.

The author refers to related work using triple-tower architectures, notably Albus et al, but he does not make clear the novel contributions of this paper. My understanding is that this is a demonstration of the architecture using teleo-reactive programs. There is also a suggestion that this architecture lends itself to learning and more on this is available on the author's web page, but I'd like to have seen a more self-contained paper here. It would only be necessary to give summary of the ideas.

I'd also be curious to know how this architecture would transfer to real robots operating in a non-deterministic world.

Although the T-R formalism is described in detail elsewhere, a brief desciption here would be helpful to understanding the example diagrammed.

A formatting note: the diagrams did not download correctly (either on Netscape or IE). I had to fetch the original Word document. Only IE correctly reproduces the symbol font characters since these are not HTML standard. So it would be better to have a postscript or PDF version.

Q2. Anonymous Referee 2 (15.10):

Recommendation: The paper can be accepted for ETAI after minor revisions. No further refereeing round is needed.

The paper is in fact excellent. Revision is only desirable in the sense of clarification or expansion of particular points.

1. Predicates corresponding to primitive percepts are associated with the "model tower". It is therefore confusing to find that the "perceptual tower" is labelled in the drawing with "Sensors".

2. A TMS is made a part of the perceptual and model towers. It is important to know whether each has its own TMS. If "yes", then is there a mediating mechanism for cases where facts stated in one of the models are left unasserted in the other, or where the two fact-bases are inconsistent? If "no", then a single unitary TMS presumably, by definition, maintains consistency everywhere and always. Whether there are two TMS's or one, questions of conflict-resolution are liable to crop up at some level, and deserve mention.

3. The paper does not make it clear whether "reporting actions" can in principle be fitted into the scheme. If so, is the reporting layer to be part of the perceptual tower? Or is it more expedient to have a separate "communications tower"?

Mutual consistency of towers

Inconsistency, mentioned above, can arise when relations qualifying one and the same object(s) enter the model tower from both perceptual and action towers. Some model assertions are updated directly from sensed environment changes. Others are derived from inferential rules of the perceptual tower. These latter, in a slightly more enriched world, might exploit knowledge of perspective etc. in handling pickup preconditions.


Suppose that an enrichment of this blocks world allows "wide", "medium" and "narrow" blocks, and corresponding hand actions set-for-wide, set-for-medium and set-for-narrow for selecting in advance an appropriate width between the grippers. Suppose that the robot naively judges the width of an object by the width of the retinal projection. Suppose also that the perceptual tower has additional information, -- whether "given", or deduced or induced, -- that although blocks normally stand front-stage, occasionally one or more starts from a position a little set back, sufficiently for a set-back medium block to subtend the same optical angle as a set-front small block etc.

The lower level of the model tower will presumably get populated with width-assertions from direct sensing, causing the robot to act accordingly, sometimes getting stuck through having the wrong gripper-width for the width of the next object to be picked up. One would like some mechanism which passes internal reports of failures to the perceptual layer, which could then perhaps plan up some action-on-failure heuristic by reasoning from everything it now knows.

A proactive perceptual tower might even intervene before the start with a proposed revision to the model tower's pick-up routine before finalizing a plan which starts by picking up x:
  if width of x is undefined then sense its width;
record the value of x's width;
set gripper-width to this value;
until pickup succeeds
pickup x;
set gripper-width wider;
record the value of gripper-width
and set to new value of x's width;
A less smart perceptual tower might have to wait for the robot to get stuck before starting attempts at cure.

Something analogous happens in human hand-eye pickup. Inconsistency can exist between the operationally perceived relative widths of two objects as implied by the "pre-grasp" posture of the hand and the epistemically perceived relative widths as measured by asking the subject to rank them. The latter is affected by the various visual illusions, the former not. In this case there are two completely separate visuo-action loops. The fast-track one doesn't use the conscious visual system at all, but a subcortical one left over from reptilian ancestry. I presume that when this primitive loop lets one down, some sort of default occurs to a more deliberative loop involving the visual cortex, among other higher centres.

These referee's comments are not to be taken as mandatory. The paper deserves to be published as it stands. The author should only respond to any of this, if at all, according to his own good judgment. The only exception might be an obligation to make the labelling of the Figure clearer.


Background: Review Protocol Pages and the ETAI

This Review Protocol Page (RPP) is a part of the webpage structure for the Electronic Transactions on Artificial Intelligence, or ETAI. The ETAI is an electronic journal that uses the Internet medium not merely for distributing the articles, but also for a novel, two-stage review procedure. The first review phase is open and allows the peer community to ask questions to the author and to create a discussion about the contribution. The second phase - called refereeing in the ETAI - is like conventional journal refereeing except that the major part of the required feedback is supposed to have occurred already in the first, review phase.

The referees make a recommendation whether the article is to be accepted or declined, as usual. The article and the discussion remain on-line regardless of whether the article was accepted or not. Additional questions and discussion after the acceptance decision are welcomed.

The Review Protocol Page is used as a working structure for the entire reviewing process. During the first (review) phase it accumulates the successive debate contributions. If the referees make specific comments about the article in the refereeing phase, then those comments are posted on the RPP as well, but without indicating the identity of the referee. (In many cases the referees may return simply an " accept" or " decline" recommendation, namely if sufficient feedback has been obtained already in the review phase).