The overall goal of this procedure is to determine how closely a computer model of a handshake matches the human handshake. This is accomplished by first presenting the subjects with pairs of handshakes, either human computer generated or a combination of the two using a tele robotic interface. The second step of the procedure is to ask each subject to choose within each pair, which of the two handshakes is more humanlike.
The final step of the procedure is to calculate from the subject's answers the human likeness grade, for the handshake model, the model human likeness grade, or MHLG. Ultimately, results can be obtained that show a significant distinction between different models, MH Gs using the Turing like handshake test. The Turing test is actually limited to verbal intelligence where scientists and engineers observed a large gap between artificial and natural intelligence in the motor control aspect of motor behavior.
Many scientists thought about this motor version of the touring test, but only recently have backed, drivable robotic devices become abandoned in research laboratories facilitating the implementation of a simple reproducible protocol for a touring like handshake test. This method can help answer key questions in the human robot interaction field, such as what are the key features of human arm movements which provide natural human touch experience. The implications of this technique extend to our diagnosis of motor impairments such as cerebral palsy.
The next step is to build a robotic diagnostic device that can distinguish a healthy human handshake from an impaired.One. Visual demonstration of this method is critical as the handshake movement through the tele robotic system is not natural and it takes time and practice to learn how to handle and interact with the robot. Following the original concept of the classical touring test, the handshake experiments consist of three entities, a human, a computer, an interrogator.
The interrogator and the human are seated in the same room but are separated by a screen, and so the interrogator cannot see which one the computer or human is generating the handshake. The handshakes are not direct handshakes, but are accomplished through a mechanical tele robotic interface made up of two phantom desktop robots in the case of a human handshake and one phantom robot and software. In the case of a computer handshake, when the word handshake appears on the computer screen, the interrogator and the human immediately make handshaking movements by moving the stylus of the tele robotic interface in the vertical direction.
The interrogator's job is to compare two handshakes and select which one is more human-like by pressing the home or F1 key to choose the first handshake or by pressing end or F two to choose the second handshake. The three experimental protocols demonstrated in this video differ from each other by the composition of the handshakes that are compared. The first protocol is the pure Turing test protocol, but the choices between a human handshake and a computer generated handshake.
The second protocol includes handshakes that are varying compositions of human generated and computer generated forces. The third protocol includes handshakes that are varying combinations of a human generated handshake and noise, or purely computer generated forces. During the experiment, a random file is created automatically at the prompt enter a name for the random file.
If the file already exists, the system uses the existing file. Otherwise, the system automatically generates a new random file with the specified name. The random file defines the order in which the different handshakes occur.
The pure protocol starts with 12 practice trials of two handshakes each in which the human and the interrogator shake hands 24 times through the tele robotic system. The interrogator knows that this stage that he shakes hands with a human. The purpose of these practice trials is to familiarize the participants with the tele robotic interface.
Following the practice round, a block of four trials is done in each trial. A human handshake is compared to one of four computer model handshakes in a random order as shown in this example. For example, in trial one of block one, the interrogator shakes hands twice.
The first handshake is generated by the human and the second handshake is generated by the computer using the forces dictated by model three. The interrogator then selects between handshake number one and handshake number two to decide which handshake is more human-like. For example, if the interrogator selects handshake number one, as more human-like the interrogator has chosen the human handshake.
However, the interrogator is not told whether he or she has correctly identified the human handshake in the next trial. The interrogator's first handshake is with the computer generated model two, and the second handshake is with the human. The interrogator again selects the most human-like handshake.
Two more trials are done in a similar manner. To test the remaining two models, model one and model four against the human handshake. This completes one block of trials.
The experiment consists of 11 blocks of four trials each with the four model handshakes presented in random order in each block. The initial block is a practice block and the analysis is done on the remaining 10 blocks. After completing the 10 blocks of the experiment, each of the four handshake models has been tested 10 times against a human handshake.
After completion of the 11 experimental blocks, a model human likeness grade is calculated for each of the four tested models. The analysis is done by calculating the proportion of handshakes in which the subject says the model handshake is more human-like than the human handshake. For example, if the interrogator chooses the model one handshake over the human handshake in two of 10 handshakes, the proportion of model to human would be 0.2, multiply this proportion by two in order to obtain the model human likeness grade.
For each model where the proportion of 0.2, the model one's model, human likeness grade would be 0.4. If model three was selected as more humanlike in four out of the 10 trials, then its proportion would be 0.4 and its model human likeness grade would be 0.8. If a model is indistinguishable from a human, its calculated proportion would be 0.5 and the model human likeness grade would be a one.
If a model is clearly non-human, like its model, human likeness grade would be zero. The MHLG is capped at one. In this protocol, the generated handshakes apart human part model where the specific combination varies across trials on a scale from a pure computer generated handshake to a pure human handshake.
Therefore, this set of experiments is referred to as the weighted human model test. The protocol begins with the practice block to help the interrogator learn to recognize a purely human handshake and become familiar with the apparatus. Therefore, the practice block has 30 trials that compare a purely human handshake to a computer generated handshake.
At the end of each trial, the interrogator chooses which of the two handshakes was most human-like. If the interrogator correctly identifies the human handshake, the screen displays correct. If the interrogator doesn't choose the correct handshake, a wrong message appears and the interrogator has the opportunity to repeat the last practice trial.
After the practice block, the interrogator begins a series of trials to evaluate two handshakes. The two handshakes are a stimulus handshake and a reference handshake, each of which is a combination of human and model forces. In the reference handshake, the handshake felt by the interrogator is an even mixture of the forces generated by the human and a computer generated model.
In other words, while the human participant makes handshaking motion, that handshake only makes up 50%of the force perceived by the interrogator with the remaining 50%of the force coming from a model in the stimulus handshake, the handshake felt by the interrogator is a varying mixture of the forces generated by the human and by a computer generated model. In some of the stimulus handshakes, the fraction contributed by the human is more than half, and thus more than in the reference handshake. In other stimulus handshakes, the fraction contributed by the human is less than half and thus less than in the reference handshake.
At the end of each trial, the interrogator is requested to choose the handshake that felt more human-like. This experiment compares three models, two test models, and one base model. As in the pure protocol, this protocol begins with one unanalyzed block followed by 10 experimental blocks.
Each experimental block consists of 24 trials comprising each of the eight linear combinations of model and human for each of the three model combinations. Upon completion of the 10 experimental blocks, a logistic psychometric function is fit to the answers of the interrogator. As described in the accompanying article, the resulting curve displays the probability of the interrogator.
To answer that, a stimulus handshake is more human-like than the reference handshake As a function of alpha stimulus minus alpha reference, the point of subjective equality or PSE is extracted from the 0.5 threshold level of the psychometric curve indicating the difference between the alpha stimulus and alpha reference for which the handshakes are perceived to be equally human-like the model. Human likeness grade for the weighted human model protocol is calculated by subtracting the PSE from 0.5, the models that are perceived as the least or the most human-like possible yield model, human likeness grade values of zero or one respectively. The final protocol compares a model handshake to a part human part random noise handshake.
In this protocol, the stimulus handshake is a computer generated handshake based on a model. The other handshake, the reference handshake is a force generated from various combinations of the human handshake and white noise with maximum absolute force of two N filtered using the Lowpass Butterworth filter of order five at a maximum frequency of two over pi. At the end of each trial, the interrogator is requested to choose the handshake that felt more human-like.
The protocol has a similar design to the previous weighted human model test. After a practice round and an unanalyzed block, a series of 10 experimental blocks is run as follows. The PSE is extracted from the psychometric curve as before and defines the model human likeness grade for the added noise protocol.
This figure shows results from the pure protocol and compares human handshakes to two different model handshakes. The two points show the calculated model human likeness grade values for two viscoelastic models named KB one and KB two. The results demonstrate that the model KB two is perceived as more human-like than model KB one.
This figure shows the results from the weighted model human protocol, which compared a varying ratio of human model handshakes to a fixed 50 50 ratio of human model handshakes. The black markers represent the model human likeness grades for the test models, and the gray markers represent those of the base model. The error bars represent the psychometric curves, confidence intervals.
This figure shows the results from the added noise protocol, which compares a model handshake to a handshake with a varying ratio of human to noise components. The black markers represent the model human likeness grades for the models and the gray markers represent those of the noise. The error bars represent the psychometric curve's confidence intervals.
These results demonstrate that the viscoelastic model KB two is perceived as more human-like than viscoelastic model KB one, using all three evaluation methods While attempting this procedure, it is important to remember to interact with the robotic device properly and try to make the interaction as smooth as possible Following this procedure. Other methods such as testing, recognition of human handshakes in pervasive developmental disorders can be performed in order to answer additional questions like, how do pervasive developmental disorders affect the ability to distinguish human from non-human gestures? Now, after its development, this technique can pave the way to use handshake as a recognition tool.
We hypothesize that the purpose of handshaking is to learn something about the other person. Therefore, the methodology we propose for developing a human-like handshake can also facilitate the development of handshake classification system, which can extract information about the other person from his or her handshake. After watching this video, it should have a good understanding of how to perform the one dimensional tour like handshake test and how we're going to evaluate handshake models submitted to the first international tour like handshake test tournament.
