Humanoid robot task recognition from movement analysis

In this paper, we present a method to perform tasks identification on a humanoid robot. The observed motion is compared to a set of candidate controllers that the robot might be executing. The more relevant candidate controllers are selected, and can be used as a description of the motion, or as a basis to replicate a similar motion on another robot. The analysis of the movements is based on the task-function approach and is applied in the context of humanoid robot motions. A pool of tasks is defined in order to cover the range of possible motion of the robot and the demonstration movement is projected in each candidate task of that pool. The reduced trajectory is then compared with the characteristic task trajectory using a numerical optimization process. The movement is then projected into the null space of the best task candidate. As a consequence, the motion due to the task candidate is subtracted from the original movement. This process is iterated until the result of a projection becomes null. The approach is applied to recognize the tasks performed by a HRP-2 robot in simulation, in order to disambiguate very similar-looking motions. Preliminary directions are given for possible application for human motion recognition.