Osaka University, Graduate School of Engineering Science, Dept. of Adaptive Machine Systems

Intelligent Robotics Laboratory -Ishiguro Lab.-

Androids

Android Science
Development of an Android
Appearance versus behavior
Generation of humanlike behavior

Android Science

Towards a new methodology of cognitive science

Both appearance and behavior are significant issues in the development of humanoid robots. However, designing the robot's appearance, especially to give it a humanoid one, was always a role of the industrial designer. To tackle the problem of appearance and behavior, two approaches are necessary: one from robotics and the other from cognitive science. The approach from robotics tries to build very humanlike robots based on knowledge from cognitive science. The approach from cognitive science uses the robot for verifying hypotheses for understanding humans. We call this cross-interdisciplinary framework android science.

A fundamental issue in android science is the existence of the uncanny valley. As robots appear more human, they seem more familiar, until a point is reached at which subtle imperfections create a sensation of strangeness or eeriness. Our important role is to verify the existence of the uncanny valley and to explore how to overcome the uncanny valley problem with androids.

Child android: the appearance is realized by making a copy of an existing person.

ReplieeQ1$B<L??(B

Adult android: it has 42 air actuators in the upper torso. The appearance is realized by making a copy of an existing person.

Facial expressions of the adult android: 13 of the 42 actuators are used in the head. Humanlike facial expressions are realized by the motion of the eyes and mouth.

Papers
Link

www.androidscience.com

By Minato

Development of an Android

Android robot - a robot that closely resembles humans -

Repliee R1

The appearance is that of a five-year-old Japanese girl. The actuator is an electrical motor. It has nine Degrees of Freedom (DoF) in the head and one DoF at the left elbow. Silicon skin covers the whole body and makes the skin feel human-like. There are four high-sensitivity tactile sensors constituted by piezo film under the skin of the left arm. These sensors can detect the touch strength. To make the appearance closely resemble a human's, the body is shaped from a mold of a girl. The skin is painted in imitation of the girl.


The appearance and internal mechanism of Repliee- R1

Repliee Q1 expo

The appearance is that of an adult woman. The face is shaped to be like an average Japanese woman. The actuator is an air-servo motor. It has thirty-one DoF in the upper body. The actuator can realize soft servo control without compliance control. An air compressor to drive the actuators can be placed apart from the android, so that the machinery noise is quiet. Silicon skin, which is the same as Repliee R1's covers her hands and part of the upper breast. There are eleven high-sensitivity tactile sensors in the upper body.

By Minato

Appearance versus behavior

Relation between motion and appearance

Using an existing robot, the effect of motion on communication was studied.
However, the appearance of a robot is also an important factor in communication. Therefore, we have to research the effect of motion in relation to appearance.
The method for researching the relation between motion and appearance is communication between androids and humans, and evaluting such communications by human psychological evaluation. The evaluations are both a quantitative evaluation (eye mark recorder) and a qualitative evaluation (SD method).

As a robot's motion or appearance comes to resemble a human's, the degree of intimacy increases. However, at a certain degree of intimacy an "uncanny valley"[Mori 1970] effect occurs. We believe that the appearance of the child type android is past the uncanny valley. However, we think that the motion of the android is near the uncanny valley. We are not sure about the adult type android, but compared to the child type android, the adult type android has more degrees of freedom. So we think the degree of intimacy is greater for the adult type android.


experimental tests

This picture shows an experiment to research the question of appearance and behavior. The participants are a dynamic android, a static android and a human. We measured the human subject's gaze quantitatively to research the influence on communication during the exchange.

The hypothesis which Mori advocated in 1970. It is called an "uncanny valley".This hypothesis says that the degree of intimacy increases with a robot's humanlike appearance and behavior. At a certain level of humanlike appearance and behavior, however, the degree of intimacy drops sharply . For example, a zombie is located near the uncanny valley of this graph.


"uncanny valley"[Mori 1970]

By Shimada

Generation of humanlike motion

It would be nearly impossible to specify manually the changing joint angles of an anrdoid with many degrees of freedom. Therefore, we explore various hierarchical methods of automatically generating humanlike motion. The first major step involves simplifying complex motions into a compact sinusoidal representation. The second involves imitating human motion as it appears at body surfaces. Past research has focused on humanoid dynamics and joint angle mappings between people and robots or avatars. However, through our experiments we are investigating what makes motion appear human. We are concerned with the structure of human motion, such as the interplay among autonomic, contingent, and conscious responses. We are also addressing how to overcome structural differences in androids and humans in order to generate humanlike motion.

Generation of humanlike motion by sinusoidal waves

Many natural-looking gestures and autonomic responses can be generated from sinusoidal waves. As a kind of "shorthand," we control the movement of each joint by specifying a sinusoidal wave or combination of sinusoidal waves. In addition, we developed the techniques of generating various periodic motions by tuning parameters of sinusoidal waves, such as their phase, frequency, and amplitude.

In an early implementation, Perlin noise determined the value of the periodic signal for each joint. While talking the android moved its arms widely, as would be typical of a political speech.

In a second example, we tuned sinusoidal parameters that determined the angles of the android's mouth, neck, torso, sholders, and arms, as shown in the following video clip:

 

Robovie II S$B<L??(B
Fig1.The example generated by Perlin noise
The value in the figure is the number of frame. ( frame is 33msec)

The generation of humanlike motion by mapping human motion to the android

We control the android by mapping human motion to its joints to accurately position the motion capture markers. However, a simple mapping from human to android joint angles is not sufficient to realize natural motion owing the android's unique skeletal structure and dynamics, which are affected by inertia and the compliance of its actuators. Therefore, a direct mapping of joint angles is not possible, and it is necessary to learn how the android's joint angles influence its body surfaces.

In this research we use the three-dimensional coordinates of markers placed on the surface of the skin for mapping from the human to the android because of the difficulty in obtaining and converting joint angles from a person to an android with differing kinematics and because of the complexity of the kinematics and dynamics of an android with many degrees of freedom. Thus, we can largely ignore the kinematic structure of the android in making the initial mapping. We can evaluate the performance of the mapping by comparing the original human motion capture data to the data generated by the android.

We use feedback error learning to control the android, and a neural network to perform the mapping of the feed-forward controller. A image of this technique is shown below:


Fig. 2. An The image of mapping motion

The following link provides a video clip of an experiment mapping human motion to the android in real time:

By Marubayashi, Matsui, and MacDorman.

Copyright © Intelligent Robotics Laboratory (directed by Prof. Hiroshi Ishigro) all rights reserved.