WAMOEBA Group

Research Overview

Self-preservation evaluation function and hormone parameters
(WAMOEBA-2 1998-present)

Rather than introducing a human psychological model established through introspection or verification using living organisms, we hoped to develop an emotional expression that would be in line with the robot’s own physical characteristics, and instead introduced the “law of self-preservation,” a basic principle of living organisms. By emotional expression, we mean emotional expressive behavior, such as blushing or breaking out into a sweat when nervous, or the bristling of a cat’s fur when it encounters a natural enemy.

Emotional Expression Robot “WAMOEBA-2”

To incorporate the principle of self-preservation into the robot, WAMOEBA-2 is equipped with the following sensors to measure its internal state:

• Battery voltage (remaining charge)
• Battery current
• Circuit temperature
• Motor temperature

In addition to movement and camera operation, a cooling fan is also installed to promote self-preservation.

Self-Preservation Evaluation Function

For each sensor used to measure self-preservation status, we defined a “self-preservation evaluation function” that evaluates the robot’s own self-preservation. The self-preservation evaluation function is a fuzzy membership function that takes the absolute value and first derivative of each sensor as input and outputs a dimensionless normalized value called the “self-preservation value.” The larger this value, the worse the self-preservation status and the more dangerous it is.

Take battery voltage as an example. A voltage value close to a full charge reduces the self-preservation evaluation value, but if it approaches the PC’s minimum operating voltage, the self-preservation evaluation value rises sharply. We also defined a voltage higher than a full charge as an overcharge, resulting in a higher evaluation value.
When combining the self-preservation evaluation values for all sensor inputs, the largest value is selected and used as the robot’s self-preservation evaluation value.

Hormone Parameters

To enable emotional expression using the robot’s body, we defined “hormone parameters.” Hormone parameters are parameters that affect the control of each part, mimicking the endocrine system of living organisms. They are defined using the absolute and differential values of the self-preservation evaluation.
WAMOEBA-2 has four hormone parameters in total. They are secreted and decay (decomposed) over time in response to increases, decreases, and absolute values of the self-preservation evaluation. We named these parameters “R-endorphin,” “R-adrenaline,” “R-noradrenaline,” and “R-dopamine,” respectively. The control parameters for each part’s degree of freedom are then adjusted according to their levels.
For example, a high level of R-endorphin will increase the pitch of the robot’s speech, and secreting R-adrenaline will increase the rotation speed of the cooling fan.

WAMOEBA-2’s Emotional Expression

The self-preservation evaluation function and hormone parameter mechanisms enable WAMOEBA to express emotions in accordance with its own physical state. For example, if the motor temperature rises during exploration behavior, the robot suppresses motor function and circulates excess current to the cooling fan. This may make WAMOEBA appear tired to an observer, and, in human or animal terms, appear to be sweating.
In this way, the WAMOEBA project demonstrated that emotional expression is possible through physical movements, even without using specific emotional parameters determined by the designer’s subjective judgment.

Adaptive Interaction Using Interactive Evolutionary Computation
(WAMOEBA-3, 2002-present)

The WAMOEBA project has focused on the theme of robots’ emotional expression based on their physical information. However, these movement parameters were previously determined by designers through trial and error.
In response to this, we introduced interactive evolutionary computation (IEC) as a method for deriving parameters appropriate for the robot’s body and the environment, including the human it communicates with. IEC refers to an evolutionary computation method, such as genetic algorithms, that uses human subjective evaluation as the fitness function.
In this research, using WAMOEBA-3 as a platform, we coded the robot’s behavioral parameters into genes, and conducted experiments in which we operated the robot using the parameters written in many different genes and evaluated its movements.

Problems with Adaptive Interaction

We found that adaptive interaction using interactive evolutionary computation (IEC) increased evaluation scores during the first few hours of learning, but then decreased them from the middle stage onward. This was due to boredom caused by physical and mental fatigue in the evaluators, leading to a shift in the evaluation axis, causing existing favorable genes to receive lower evaluations.

Reducing Evaluator Fatigue through Hybrid Evaluation

In evolutionary computation, maintaining diversity within the gene pool is considered effective in responding to dynamically changing evaluations. Possible ways to maintain diversity within the gene pool are to increase the number of individuals or the mutation rate. However, in IEC, increasing the number of individuals increases the number of evaluations required by the evaluators, which increases their physical fatigue. Furthermore, increasing the mutation rate reduces the learning convergence rate, which may also lead to subject boredom.
In contrast, in this study, genes were analyzed and discriminated using a self-organizing map, which suppressed the evaluation of similar genes and reduced fatigue among the evaluators, while also increasing the number of individuals in the gene pool, thereby successfully maintaining high evaluation values in the middle of the experiment.