Parametric Hidden Markov Models for Recognition and Synthesis of Movements

In humanoid robotics, the recognition and synthesis of parametric movements plays an extraordinary role for robot human interaction. Such a parametric movement is a movement of a particular type (semantic), for example, similar pointing movements performed at different table-top positions. For understanding the whole meaning of a movement of a human, the recognition of its type, likewise its parameterization are important. Only both together convey the whole meaning. Vice versa, for mimicry, the synthesis of movements for the motor control of a robot needs to be parameterized, e.g., by the relative position a grasping action is performed at. For both cases, synthesis and recognition, only parametric approaches are meaningful as it is not feasible to store, or acquire all possible trajectories. In this paper, we use hidden Markov models (HMMs) extended in an exemplar-based parametric way (PHMM) to represent parametric movements. As HMMs are generative, they are well suited for synthesis as well as for recognition. Synthesis and recognition are carried out through interpolation of exemplar movements to generalize over the parameterization of a movement class. In the evaluation of the approach we concentrate on a systematical validation for two parametric movements, grasping and pointing. Even though the movements are very similar in appearance our approach is able to distinguish the two movement types reasonable well. In further experiments, we show the applicability for online recognition based on very noisy 3D tracking data. The use of a parametric representation of movements is shown in a robot demo, where a robot removes objects from a table as demonstrated by an advisor. The synthesis for motor control is performed for arbitrary table-top positions.