Reading through some of the articles linked to in the Wikipedia article, I'll respectfully disagree with @Theran. The distinction seems quite well grounded, although Wikipedia does a poor job of making it.
The term embedded systems (ES) has been around since the 60s and can, arguably, refer to anything from an airplane to a Furby. I think the term cyber-physical systems (CPS) was coined to distinguish it from what are traditionally thought of as embedded systems, namely closed-loop, non-networked "boxes" that operate in a very well-defined and constrained domain with a limited power to affect physical systems. CPS, on the other hand, embody the idea of think globally, act locally (my apologies to Patrick Geddes), that is, they are usually highly networked systems that bring about change in a local physical system dependent on the state and actions of other entities in the wider network.
While many robotic applications fit this definition, and can therefore be termed cyber-physical systems, many are not. What bestows the honour on MIT's robotic garden, I believe, is the fact that the robots form part of a wider, decentralised system (PDF). It is the plants, equipped with sensors, that decide when to request watering or other services from the robots, while it is the robots that then decide between them which one will fulfil that request. Furthermore, not all CPS are thought of as "robotic", for example, an intelligent power grid.
Cybernetics, as @Theran has noted, is occupied with the study of control systems, and so will form a core part of studying CPS, but also has a broader range of applications in fields such as mathematics, economics, and sociology, for example.
This report on cyber-physical systems (PDF), by Edward Lee from UC Berkeley, makes clear that CPS are a next step in the evolution of embedded systems with many of the same constraints (real-time capabilities, reliability) plus a few extra ones (robustness, adaptability, intelligence, interconnectedness). As such, the field of CPS is, in parts, concerned with developing completely new approaches to hard- and software architecture. For example:
But I believe that to realize its full potential, CPS systems will require fundamentally new technologies [...] One approach that is very much a bottom-up approach is to modify computer architectures to deliver precision timing [...] Complementing bottom-up approaches are top-down solutions that center on the concept of model-based design [...] In this approach, "programs" are replaced by "models" that represent system behaviors of interest. Software is synthesized from the models.
Lee's thoughts are echoed in this Embedded Computing column (PDF) by Wayne Wolf of Georgia Tech.
After all, we've had computers attached to stuff for a long time. Why, you may ask, do we need a new term to describe what we've been doing for years? [...] We have a surprisingly small amount of theory to tell us how to design computer-based control systems. Cyber-physical systems theory attempts to correct this deficiency. [...] Cyber-physical systems actively engage with the real world in real time and expend real energy. This requires a new understanding of computing as a physical act—a big change for computing.
I recommend reading both articles for a good view on how CPS are different from "mere" embedded systems. Cyberphysicalsystems.org has a concept map of CPS on their homepage that nicely illustrates many of the aspects involved in developing CPS.
As for the origin of the term, none of the sources I found attributed it to anyone. Many papers defined it without attribution while clearly not being the first to use them. The term first crops up in the literature in 2006 but, by that time, the US National Science Foundation had already organised a Workshop on Cyber-Physical Systems, suggesting that the term was already in use by then.