Here are some of my thoughts on this. Im implmeenting something very similar for the game im working on. The AI ive built uses a similar scheme to map its goal->action->act states. Good Luck!

The issue seems to be a mapping from a set given commands to an arbitary effector object whose capablitites are not known at compile time.

To simplify the design it would be nessecary to create a method in which the higher controlling module can introspect itself to determine what possible actions it can perform.

By classifying actions, a designer/AI architect can build AIs at a higher level and plug and play them in any body design, however there is a draw back in that, it''s possible for modules to implement ablities which the higher level AI would have no concept of, such as teleporting or burrowing.

What''s needed, perhaps is another level of abstraction within the high level controller module, between intention and actions. If intention can be defined, and a subset of possible actions an AI body can perofom eumerated. Then it would be possible to do a search through that space to find an adequate solution if one exists or if CPU cycles are available.

For the AI to make a somewhat intellegent choice between the set of actions, it must be able to estimate the actions cost relative to other actions. The cost computation is context senstive.

Intention are goals the AI sets for itself. The AI will exhibit behaviors as rich as this statespace will allow. However there is no guarentee that a large state space will make a richer AI character, since some states might be redundant or unreachable.
Goals can be move_to, attack, help_another, etc.. Actions can be take_stepA->B,burrowA->B,turn_A_degrees,face_another, etc..

They are similar, i suppose their only difference is granality represenation.

Good Luck!

-ddn