Hi @vclay
I did have brief discussion with @tslominski about the feasibility of using Monty. Tristan explained that TBP is focussing on cortical (and cortical adjacent) brain structures rather than the cerebellum and brain stem structures involved in fine motor control and muscle memory. So at present Monty would only go as far as generating a signal to walk, and not contain the pattern generation capability required by robotic limbs to walk, what I have referred to as a top-down approach in my documents. I have been pursuing a bottom-up approach in the hope of learning about how to implement higher functions by implementing lower level functions first.
I could talk at length about the brain structure I have in mind, but here is precis 
I plan to build the brain using artificial neurons (weights, bias, activation function etc) but connected in specific, initially handcrafted, structures rather than the regular arrays commonly used in AI.
One group of neurons will create a state machine to move the limb from one position to another position. An output group of neurons will convert the generic ‘move’ signals to muscle-memory specific positions stored in the output neurons (the muscle-memory will be specific to each robot unit and will have to be learned and updated over time).
The muscle-memory signals will go to the servos and also be fed back to a simulation group of neurons. The simulation neurons, with the addition of a temporal input, will generate a linear simulation of the expected motion of the limb motors. The outputs from the simulation will be fed back to the input layer.
The input layer will compare expected motion against active motion, any significant deviation will activate a connection to alert the state machine and higher level control to a possible problem. The input layer also compares actual position against final target position to trigger the state machine onto the next state.
All of this structure is replicated for power usage, setting maximum power to use in the move and checking to ensure it is not exceeded. Speed of the move is another variable in the mix, as is foot pressure.
A very simple diagram is shown below, where each block represents many neurons. Neurons will have multiple local connections and some long distance connections much as we see in a biological brain. For example, an over-power signal, the robot equivalent of pain, will go to the local state machine group to stop the move immediately and also to many other regions to alert the whole brain to the situation, just as when we step on a sharp object our whole brain is alerted to the event. Additional blocks will handle strategies to deal with problems such as the foot not hitting the ground when expected.
I intend to implement the neurons and the connections between them as a large data structure. Each ‘record’ in the data structure will represent a neuron, its activation function, weights and connections from other neurons. The data structure will expand to gigabytes and eventually terabytes. Processing of the data structure can be carried out by whatever means are available, be it software, GPU or even arrays of FPGAs. The data structure and all the knowledge it contains remains independent of the means of execution and thus hopefully protected again obsolescence.
The downside is that none of the tools required to create these structures currently exist, as far as I am aware, and so they also need to be created alongside the brain structure itself. Whilst the initial networks can be hand-crafted it will soon reach the point where the structures need to be defined at a higher level and compiled by a new type of neural network compiler.
I will try to create a video showing a walking motion as depicted by nerve signal patterns coming from a single limb, the melody of motion I referred to some time ago.
In answer to the obvious question ‘how does it learn’ I can only say, I’m working on that 
, but it’s not back propagation from a global error signal, no such signal exists. Rather it will be at a group level to correct mismatches between simulation and sensor input.
An important point is that the systems is always active, sensory data is always pouring in and the neural network is continually responding to any changes.
) and maybe some discussion around the brain part of it!
