Dynamic motion adaptation for 3D acrobatic humanoids

Low-cost dynamics is a key issue in planning complex motions. This paper aims at proposing a fast algorithm in order to simulate aerial motions for humanoids while using motion capture data. As the real subject has obviously different anthropometric parameters than the synthetic robot, directly applying motion capture data leads to a non-respect of physical laws. As a consequence, the humanoid can land in a bad configuration. During aerial phase, the acceleration of the center of mass may be different from gravity and the angular momentum may change incorrectly. This so-called retargetting problem is generally addressed in computer animation by solving kinematic constraints. In this paper, we propose a dynamic filter that is able to rapidly adapt motion in order to preserve its physical correctness. If the original motion is changed by a user or just replayed on a different skeleton, the filter is about to correct the postures at each time step in less than 2 ms in a common computer. It's consequently possible to change an arm motion during the aerial phase and to immediately check its effect on the global rotation of the humanoid.