| Summary: | 碩士 === 國立臺灣科技大學 === 資訊工程系 === 102 === Humanoid Robots, especially small-sized ones, are considered mature products theses day. Small-sized humanoids are adequate for teaching and entertainment purposes. However, the degrees of freedom (DOF) of small-sized humanoids are usually less than life-sized humanoids, not to mention human. Thus, it becomes an issue when we want small-sized humanoids to perform human-like smooth motions by their insufficient numbers of DOF.
This study focuses on issues about translating human motions to humanoid commands, including motion similarity improvement and lower body balancing. There are three parts in this study. For arm motions, naive translation usually generates different wrist positions due to insufficient numbers of DOF and the movable range limit in each motor. The first part discusses a similarity measure based on the positions of elbow and wrist such that we can fine tune the generated arm motions. The second part explains how the absence of waist degree may destruct motion similarity, and how to compensate this problem on small-sized humanoids when waist rotation is not available. Finally, the third part introduces the differences between the structure of human legs and humanoid legs. A few tricks are then developed to adjust hip and heel joints to keep the humanoid robots from falling when performing a translated motion.
To verify the above-mentioned techniques, we adopted the CMU motion capture database as input and performed the improved translation on nine selected motions. These motions are then played back on MotionBuilder, a low-end humanoid, and observed by 13 people. The results of feedback show that fine tuned motions obtain higher scores than direct translation version for all 9 motions.
|
|---|
