This model represents action as a continuous transition of joint rotations between poses. Each joint is modelled separately and can have changes in direction. All movement is synchronized at poses.
Each between-pose movement is represented by separate data. It is fully independent except for the overall duration, which is controlled by the sequence of poses.
To get a feel for this representation, I suggest that you try programming some movements. The rules are simple. The movement goes through different values in the joint angle. The values are entered, separated by semicolons. Each value is associated with a 'keyTimes' entry. These entries are also separated by semicolons. The first 'keyTime' is zero, representing the start of the movement (the value for the initial pose). The last 'keyTime' is 1, representing the end of the movement (the value for the terminating pose). The entries between represent the fractions of the duration between poses. 'keyTimes' entries are also separated by semicolons. For this experimentation, the duration is set as 5 seconds.