Body Language for Personal Robot Arm Assistant

Abstract

This article presents an exploratory study aimed at: (i) understanding participants’ first impression of the Cyton Gamma 1500, a small machine-like robot arm, and (ii) designing robot body language to convey certain intentions with regard to a personal assistant robotic application. The between-group study comprised of three distinct groups of participants. Each group had a different first encounter with the robot arm: the first group saw an idly sitting robot arm; the second group watched another person working with the robot; lastly, the third group worked with the robot themselves. They rated the robot arm on a bipolar adjective questionnaire after the first encounter which gives us insights into participant’s “first impression”. Then they performed a categorization task where they chose the most suitable posture for conveying specific messages. The objective here was to select postures that can easily convey the robot assistant’s intentions to users in future. The two main findings of the study are: (1) participants’ impression of the robot arm does not differ significantly across the three groups despite having different “first encounter” with the robot, and (2) certain postures are easily relatable to the message they convey (like, Robot is saying “Hi!”), while certain others are not (like, Robot has surrendered). In light of phenomenological theories of communication, it was concluded that tempered forms of anthropomorphic or zoomorphic postures are more easily identified than completely abstract postures. The findings of this study can be used to design robot-specific body language for robot arms.

Appendices

Appendix A

All postures used in the categorization task.

Appendix B

See Appendix Table 4.

Table 4 Results of the categorization task in all three conditions

Appendix C

Statistical analysis using Shannon entropy and the Hellinger coefficient.

The Hellinger integral was introduced by Ernst Hellinger in 1909. This coefficient is used to quantify the similarity between two distributions. For two discrete probability distributions, $P=({p}_{1},\dots ,{p}_{k})$ and $Q=({q}_{1},\dots ,{q}_{k})$, their Hellinger distance is defined as

$$Hel(P,Q)=\sqrt{\frac{1}{2}\sum_{i=1}^{k}(\sqrt{{p}_{i}}-\sqrt{{q}_{i}}{)}^{2}}$$

In the present context of a categorization task, we use the Hellinger’s distance coefficient as an indication of the distance between the observed distribution and the uniform distribution, where the observed distribution refers to participants’ choices regarding a posture across the eight statements. The larger the Hellinger’s distance, the greater the difference between the observed distribution and the uniform distribution.

The concept of Shannon entropy finds more usage in information theory, where the message sent by a transmitter is modified when it passes through the channel, and the receiver attempts to infer the message that was sent. Shannon entropy gives the expected value (average) of the information contained in each message. “Messages” can be modelled by any flow of information. Shannon entropy was used to obtain a numeric value of the strength of appeal of each posture. The lesser the value of Shannon’s entropy, the clearer the message communicated by the posture.

Shannon defined the entropy, $H$ (Greek capital letter eta), of a discrete probability distribution, $P=({p}_{1},\dots ,{p}_{k})$, as

$$H(P)=-\sum_{i=1}^{k}{p}_{i}{\mathrm{log}}_{b}({p}_{i})$$

where b is the base of the logarithm used. Common values of b are 2, Euler’s number, e, and 10. The unit of entropy is Shannon for $b=2$, nat for $b=e$, and hartley for $b=10$. In our case $k=8$ (8 statements), and so we chose to work with $b=8$ so that $H$ is bounded within the [0, 1] interval. In the case of no information, where ${p}_{i}=1/8$—a uniform distribution—$H(P)$ will be 1, and in the case of full information, where ${p}_{i}=1$ for a single value of i and zero otherwise, $H(P)$ will be 0.