Ideally, from a hardware point of view, you are probably going to want to obtain some kind of robot chassis; for development and experimentation.  Since Elon's robots aren't shipping yet; nor am I aware of an affordable variant of some of the really cool stuff that Boston Dynamics is doing; we will probably have to settle for a much simpler platform such as an upgraded Parallax Boe-Bot; which can be converted to work for example with a Atmega 2560 board and a Parallax prototyping board, as seen in the pictures.   Thus from a hardware point of view, this is a pretty simple project, since you can just as easily get rid of the robot altogether; and just use any Arduino or Parallax prototyping system, which is capable of "blinking" one LED or another when we give it commands like "KiLL TROLL WITH SWORD!", or "FIND CRYSTAL PYRAMID IN MAZE."  Eventually, that is the sort of thing that we should hope to accomplish; whether in the real world or in some simulated universe.  Yet for now - because this project is tickling with the forefront of AI, it should be understood that this is going to be a VERY hard project from the software standpoint, yet the efforts spent, will hopefully be well worth it.

Thus one approach might be to obtain the latest C++ version of Mega-Hal which I have put in the Git-Hub repositories, along with some additional tools; which can be used to create a "make-file-like" system for creating training sets for pretty much any AI that can be trained on standard text input.  Likewise, there is a port of the UCSD Pascal compiler that is "mostly complete" which is also in the Git-Hub repositories; which will eventually turn out to be useful, not only for creating new build tools for microcontrollers; but also for improving how AI deals with "logical problems" in general; i.e., as far as handling problems of "All men are mortal" and so on; by allowing the AI to handle class hierarchies and reflection at runtime.  The Propeller Debug Terminal source in turn provides a method for "getting out the the manufacturer's sandbox" as it were; so that you can interact with your bespoke applications in a more highly customizable way than the usual vendor-provided tools.

As can be seen from the project logs, I have achieved some "interesting" results, by training an AI on some principles of compiler design and music transcription theory; and then I managed to get it to develop an interest in religion, sex, and of course a "claim to have an outline for the Yang-Mills conjecture" of its own, that is when it is not possibly "streaming to the galaxy", or whatever it actually said.  As the size of the training set increases; the AI should eventually be able to handle "Zork-like" commands on a real physical robot, as discussed; even though that is going to be a VERY hard project, it is easy to see, based on what has been learned about LLM"s for example, in recent times; that this is very doable - even with modest hardware.  So if you have ever wanted to build a real-life SHRDLU system with an actual robot hand that can pick up things and do things in response to typed or spoken commands; then this project is the "missing piece" to that puzzle that will help you bring all of that other hardware fun together.

Project integration across multiple domains therefore is the real purpose of this project.  Enjoy!