* The pose argument is the local link pose.
  * The origin of either visual or collision node can be used.

  * Each body can be made from multiple meshes with their own set of
    vertex indexes.
  * When adding faces of a new submesh, offset the face indexes by the
    number of points already added to the polyhedron.
  * Another cleaner solution would be to load assemblies of
    polyhedrons, but it would not comply with robotbuilder behavior
    for HRP-2 (one body = one polyhedron).

  * Find hpp parent joint by name and get absolute position.
  * Denormalize joint to get urdf joint absolute position.
  * Multiply by link relative transformation to get absolute
    transformation.

 * Joint normalization was called recursively for all branch joints
   when computing the position of the current joint.
 * Call joint normalization only on current joint and before applying
   transformation.

  * Create appropriate Kite solid component and attach it to body.
  * Use assimp to parse mesh files.
  * assimp usage has been copied from rviz and adapted.
  * Only meshes are supported for now.

  * The methods are largely inspired from the similar parser
    jrl_dynamics_parser by Florent Lamiraux, Guido Manfredi, and
    Thomas Moulard.
  * So far the dyncamic part of the robot is fully loaded, as the
    kinematic part in of Kite device component.
  * No Kite solid component is attached to the bodies.