Quantification of Human-Like Motion of Androids in Relation to Control Parameters

In android science, the greater the resemblance between an android and a human, the greater is the expectation of observers for its movements to appear natural; therefore, the "human-likeness" of android motion is important. Traditionally, studies on the human-likeness of motion have addressed two separate problems: generating motion trajectories and realizing the target trajectories on hardware. However, trajectory generation often relies on subjective evaluations and lacks quantitative metrics, whereas trajectory realization has limited reproducibility owing to the physical characteristics and control performance of the android. Therefore, treating these problems separately complicates the task of achieving human-like motion. This study proposes an approach that quantitatively addresses the human-likeness of such movements under the assumption that the target trajectory can be stably reproduced. As a control method, we adopted Active Disturbance Rejection Control (ADRC), which can suppress disturbances and produce stable responses even for complex nonlinear systems. We aim to express the human-likeness of movements realized under this control by using engineering metrics. Specifically, we computed the engineering metrics from motion data obtained by varying theADRCparameters and quantified the subjective perceptions of human-likeness based on pairwise comparison experiments. The analysis of the relationship between these two sets of measures confirmed a reasonable agreement between the ranking trends of the engineering metrics and the subjective evaluations. These results indicate that the proposed engineering metrics can quantitatively describe human-like motion independent of subjective assessments, and are effective as objective and unified indicators for android motion design.