2014 Fiscal Year Annual Research Report
Development of a humanoid robot motion teaching system based on the interpretation of tactile instructions
| Project/Area Number |
26880014
|
|---|
| Research Institution | Osaka University |
|---|
Principal Investigator |
ダーラリベラ ファビオ 大阪大学, 基礎工学研究科, 特任助教 (80740092)
|
|---|
| Project Period (FY) |
2014-08-29 – 2016-03-31
|
| Keywords | 人間型ロボット / 接触 |
|---|
| Outline of Annual Research Achievements |
Touch has the possibility of allowing inexperienced people to intuitively teach motions to humanoid robots. However, it is a complex matter to interpret the meaning of tactile instruction. Preliminary experiments showed that when people are free to associate meanings to tactile instructions, different people give different meanings to the same instruction.
The purpose of this research is to identify how robots should interpret touch instructions and advance the state of the art of robot motion development systems.
To do this, during this year an extensive amount of associations between touches and movements (over five thousand) were collected from users that interacted with the robot for the first time. In order to identify what types of responses people expect, unsupervised clustering was applied to the collected data.
The proposed clustering methodology was shown to be able to capture people's common sense: it is able not only to decide whether two expected responses are similar, with good agreement with people's qualitative judgement, but also to identify whether a response is natural or not, by computing the probability of belonging to one cluster.
This similarity between the automatic clustering and people's common sense provides support to the idea that the automatically identified types of responses correspond to a plausible classification of robot's responses to touch. The results were reported in a paper titled "How do people expect humanoids to respond to touch?" submitted to the International Journal of Social Robotics and currently under review.
|
| Current Status of Research Progress |
Current Status of Research Progress
1: Research has progressed more than it was originally planned.
Reason
As was initially planned, the association between tactile instruction and expected movements was studied by asking subjects the response they expect from the robot when it is touched. The initially planned classification was obtained by using unsupervised clustering. Several clustering approaches were compared, and the results showed that a Naive Bayes Model is very suited for capturing similarities and differences among touch responses. An initially unforeseen property of the classification algorithm was identified: the maximum probability of a response to belong to a cluster is an index of the naturalness of the response. This experimentally confirmed fact can be explained by observing that if responses are close to the prototypical responses, and hence they have high probability to belong to a cluster, then they are likely to be natural, while if they are far from any prototypical response they may be difficult to interpret both by the clustering algorithm and by people.
Analyzing people's expectations in the field of tactile communication is expected to be beneficial in many areas. For instance, during our data collection, many people expected that when the robot is touched somewhere, it would watch the touched part. This kind of reactions may be used as an early indication of accidental touch for increasing safety in the interaction. Feedback helps users in not loosing interest in teaching, and motivates them in putting more effort in the process. Many of the responses identified may be used as feedback in contexts in which users touch the robot for teaching.
|
| Strategy for Future Research Activity |
In the remaining period, the research will be conducted following the initial plan. In detail, focus will be given on how touch can be used to teach motions to humanoid robots. Interpreting the meaning underlying touch instructions is complex. Different people tend to give different meanings to the same instruction, and even the same person may associate different meanings to the same touch, depending on the context.
Research will focus on clarifying whether there are intuitive protocols that, once learnt, allow people to teach robots efficiently. Three hypotheses can be made: First, that there is one optimal protocol, efficient for everybody. Second, that there is a set of optimal protocols; each person finds some protocols intuitive and efficient, and cannot develop motions easily with other protocols, even though those are the most intuitive for other people. Finally, the third hypothesis is that only the protocols each user defines by oneself are intuitive and thus efficient.
If the first hypothesis is verified, then an optimal fixed teaching system will be developed. If the second hypothesis is true, in the developed teaching system the user will be able to choose among a fixed set of possible protocols. Finally, if the third hypothesis is true, it will be necessary to develop a system that adapts to every single user.
In each of these cases, we expect to obtain a highly efficient system that lets inexperienced users teach motions to humanoid robots.
|
