Here is a somewhat exotic suggestion, different than Google glasses, which I admit might not fit your program.
Basically, science is augmented reality. Every time Mathematica is used to produce some model of reality, it contributes to augmented reality. So what could be done ?
Schematically, science begins with observations or experiments, then continues with assumptions and reasonings that lead to models or theories. That is where Mathematica commonly intervenes by computing representations of reality that scientists compare with reality with the purpose to adjust and improve their models or theories. Then, when the models are relevant, they are used to steer the design or creation of technical objects or possibly to drive instruments or processes.
From this unusual viewpoint about augmented reality, Mathematica could be involved in the experimental process by some capacities to connect to scientific instruments (or the files they generate) and directly get the data. It could also be involved in the process of producing models or theories, typically by including tools like system modeler. Then, the models can be superposed on the data, either to test whether they fit them or to merge the data with the models, which constitutes augmented reality.
There is a symmetric application in which Mathematica might be involved too, when the computer is used to reconstitute a more or less realistic simulation of reality from some model, which is commonly called virtual reality. Mathematica might be more involved in this domain thanks to, beyond its current solvers, future finite element solvers or other kinds of solvers, a discrete-time version of Animate in view of stepwise simulations (a stream approach, as opposed to ListAnimate) and possibly with sorts of high level dynamic graphic primitives able to implement scenarios.
I believe there are three complementary domains in which Mathematica might also contribute to augmented reality.
The first one regards the modeling process, which involves not only computations but also the writing, reading and analysis of scientific publications. A future version of Mathematica could manage warehouses of "hyperlinked" publications and data, probably with the help of CDF and WolframAlpha plus a Google approach to the Mathematica notebooks put on the web; a sort of Mathematica web, better structured than the current awkward html-xml-php-javascript-sql web.
The second one regards the Mathematica language which is transformational, while the scientific language is more relational, so a logical extension of Mathematica would be welcome. Actually, Roman Meader developed such a logical extension on top of Mathematica, a sort of little Prolog designed in the frame of the Mathematica syntax. Unfortunately, this package no longer works (since version 6); an updated version would be welcome. Besides, I believe there could be interesting couplings between logical programming, graph algorithms and database design, but this is probably a topic for another discussion thread
The third one regards the design of objects or the driving of processes, for which an extended version of Controller* might enable these sorts of operations directly from within Mathematica. 3D printing is a kind of first step in this direction but the link should be thought not only with devices but also with specialized pieces of software.
In all cases, there should be a theory of that, I mean a formulation able to describe such features and above all to discuss and reason about them.