Help me to make something more natural than "set-joint-velocity"

I'm trying to rig up some multibodies with specific configurations of joints and links in order to start making some experiment with cesarg's excellent python implementation of NEAT in order to evolve neural net controllers for virtual creatures. As an example of the kind of behaviour you can see with this, check out http://ero.matfyz.cz/ which already incorporates NEAT with ODE (the same physics engine breve uses). Although they have demos/videos posted, it looks like they are currently doing work on that right now so I guess we shouldn't hold our breath on them sharing any of it. The important shared characteristic between those simulations and the brevecreatures experiment packed in breve is that the neural nets (or sin transformation parameters in the case of brevecreatures) interface with the physical joints by directly applying a force.

This is fine when your simulation is made up of idealized abstract joints which connect tangible physical objects; except that it produces unnatural movement!
If you look at biological joint/muscle interactions they are easily broken down into four subtypes: holding, relaxing, flexing and extending.
This is because in real life, the nervous system of an organism is able to manipulate the position of joints by controlling opposing groups of muscles. The action of "holding" a joint is actually the action of applying balanced forces to the opposing muscle groups in order to achieve a stable position. Compare relaxing, which is applying no force to either muscle group; extending, which is tensing only the outer muscle group; and contracting, which is tensing only the inner muscle group.

There are good emergent properties of the opposing muscle joint morphology: the muscles provide can elastic force when the joint is moved far from its "natural" position; this elasticity can be leveraged by alternatively relaxing and tensing a joint in order to provide far greater levels of peak torque along that joint.

If you were simply providing set-joint-velocity or set-joint-torque to your evolved controllers, they will have to accomplish the act of "holding" a joint position by oscillating violently from one direction of max force to the other. Also, if a creature is in a situation where it has to constantly be applying force, its impossible to take advantage of joint's natural abilities to relax. This is not a natural way to move and I think there is probably a way to do better.

If you want to make natural and effective locomotion in breve, you should start by adding another layer to the creature/brain interface: the creature should be choosing an action for each joint first (i.e. flex, extend, relax, or hold), and only THEN should it consider the relevant parameter for that action (for example the force with which to extend, contract, or oppose movement).

I'm completely confused as to how this should be implemented to fit into an experiment. Does anyone have any thoughts on what I have to do to make this?