If you generate the movement procedurally using physics, then you do not need animation data at all.
As soon as you start to incorporate animation data, the movements will be 'canned.'
If you want to see what the real deal looks like:
(videos)
Bram
Not true. In fact, in the videos you posted they -do- use pose information; in video 2 of 2 near the end they explain:
"A motion generator outputs target poses throughout the walk cycle and is used to produce different styles of motion."
Target Poses ARE 'animation' as you put it. That's what bare mettle means when they say they want an animator -- to provide more realistic target poses for their procedural physics system to animate through. The target poses give the posture of the characters, but they react to the environment through all their joints just like the video you linked.
A simpler version of this style is the same as how games like Dark Souls and Dragon's Dogma allow their characters' feet to never collide through stair steps etc, even when using animation targets. The animations are used first, but passed through additional functions that predict the world environment and clamp movement or apply new rotations to joints. That way, if a foot is going through a default animation heading towards a stair step, the physics engine will pass data through the character's skeleton telling the bone joints to stop moving when they have encountered an obstacle.
Source: Game design school, Unity Developer, & Game Developer's Conference 2012 - Animation Seminar
P.S.
Something those videos do bring up, however, may be a solution to the 'drunk' feet movements of the SG characters. 'Velocity tuning' may be what's necessary to help correct the 'drunk' or 'swimming' leg movement.