行政院國家科學委員會專題研究計畫成果報告
計畫編號：NSC
96-2221-E-110-081-執行期限：96 年 8 月 1 日至 97 年 7 月 31 日
主持人：李偉柏
執行機構及單位名稱：中山大學資管系
計畫參與人員：楊宗憲 中山大學資管所 博士生
蕭羽廷、鍾佶修 中山大學資管所 碩士生
Abstract Designing robots for home entertainment has become an important application of intelligent autonomous robot. Yet, robot design takes considerable amount of time and the short life cycle of toy-type robots with fixed prototypes and repetitive behaviors is in fact disadvantageous. Therefore, it is important to develop a framework of robot configuration so that the user can always change the characteristics of his pet robot easily. In this project, we present a user-centered interactive framework that employs a neural network-based approach to construct behavior primitives and behavior arbitrators for robots. For evaluation, we use the proposed framework to construct emotion-based pet robots. Experimental results show the efficiency of the proposed approach. Keywords: user-centered design; pet robot; artificial evolution; emotion modeling; neural network; behavior learning; behavior arbitration 1. Introduction
In recent years, designing robots for home entertainment has become an important application of intelligent autonomous robot. One special application of robot entertainment is the development of pet-type robots and they have been considered the main trend of the next-generation electronic toys [1]. This is a practical step of expanding robot market from traditional industrial environments toward homes and offices.
There have been many toy-type pet robots available on the market, such as Tiger Electronics’ Furby, SONY’s AIBO, Tomy’s i-SOBOT and so on. In most cases, the robots have fixed prototypes and features. With these limitations, their life cycle is thus short, as the owner of pet robots may soon feel bored and no
longer interested in their robots. In addition, every individual user expects a personalized robot that has unique characteristics and behaviors, and he can train and interact with the robot in a specific way. Sony’s robot dog AIBO and humanoid robot QRIO are sophisticated pet robots with remarkable motion ability generated from many flexible joints. But these robots are too expensive to be popular. Also the owners are not allowed to reconfigure the original design. Therefore, it would be a great progress to have a framework for robot configuration so that the user can always change the characteristics of his robot according to his personal preferences to create a new and unique one.
Regarding the design of pet robots, there are three major issues to be considered. The first issue is to construct an appropriate control architecture by which the robots can perform coherent behaviors. The second issue is to deal with human-robot interactions in which natural ways for interacting with pet robots must be developed [2]. And the third issue to be considered is emotion, an important drive for a pet to present certain level of intelligence [3]. In fact, Damasio has suggested that efficient decision-making largely depends on the underlying mechanism of emotions, and his research has shown that even in simple decision-making processes, emotions are vital in reaching appropriate results [4]. By including an emotional model, the pet robot can explicitly express its internal conditions through the external behaviors, as the real living creature does. On the other hand, the owner can understand the need and the status of his pet robot to then make appropriate interactions with it.
To tackle the above problems, in this project we propose an interactive framework by which
the user can conveniently design (and re-design) his personal pet robot according to his preferences. In our framework, we adopt the behavior-based architecture ([5]) to implement control system for a pet robot. Different interfaces are constructed to support various human-robot interactions, including device-based control, speech-based control and gesture-based control. To evaluate our framework, we use it to construct a control system for the popular LEGO Mindstorms robot. Experimental results show that the proposed framework can efficiently and rapidly configure a control system for a pet robot, and its human partner can interact with the pet in different ways. In addition, experiments are conducted in which a neural network is used for the pet robot to learn how to coordinate different behaviors for a user-specified emotion model. 2. Metholology
Our aim is to develop a user-centered framework that can assist a user to rapidly design a pet robot. Robot design involves the configuration of hardware and software. Expecting an ordinary user to build a robot from a set of electronic components is in fact not practical. Therefore, in this work we concentrate on how to imbue a robot with a reliable and robust control system.
Our framework mainly includes three modules to deal with the problems in building pet robots. The first module is to design an efficient control architecture that is responsible for organizing and operating the internal control flow of the robot. Because behavior-based control architecture has been widely used to construct many robots acting in the real world successfully, our work adopts this kind of architecture for robots. The second module is about human-robot interaction. As a pet robot is designed to accompany and entertain its human partner in everyday life, interactions between the owner and his pet are essential. Our framework provides two natural ways, voice-based and gesture-based methods, for human-robot interaction and communication. The third and the most important module is a learning mechanism that includes two parts to resolve the problems of behavior creation and behavior arbitration. The first part is to evolve/learn new behavior primitives; and the second part, a neural network-based emotion system for behavior arbitration. With the emotion
system, a pet robot can act autonomously. It can choose whether to follow the owner’s instructions, according to its internal emotions and body states. Our framework is designed to be user-centered and has a modular structure. The user can use it build and change any part of the control system for his robot.
To automate the process of building behavior primitives, we use an evolutionary approach to learn behavior primitives. In this work, we employ a genetic algorithm system to evolve reactive behavior controllers, in which a controller is a two-layer feedforward neural network. The encoding scheme here is a one-to-one mapping in which each weight and bias of the neural network is a floating point number corresponding to a gene within the string chromosome.
To coordinate different behavior controllers obtained from the above procedures, our framework uses a special mechanism that exploits emotions for selection of behavior. To implement the emotion model, we need to select the emotion elements to constitute the model. As our goal here is to establish a framework for constructing personal pet robots, rather than to investigate the interactions and relationships between cognitive and emotion systems, our work only models basic emotions to coordinate the behave controllers the user has pre-chosen for his pet. To quantify the emotions, we define each of the basic emotion as a special mathematical formula. In addition to emotions, another set of internal variables is defined to describe a robot’sbody states.
The emotion model determines the innate characteristics of a robot, which are highly related to its abilities of learning and adaptation. As we have emphasized, the design of pet robot must be user-centered. Therefore, in addition to the emotion model described above, the user’s opinions (or expectations) need to be taken into account in training a pet robot. To achieve this goal, the neural network module is used to map a set of emotion and body state data into appropriate behaviors under the user’s guidance. 3. Results
To evaluate the proposed ANN-based methodology, experiments have been conducted. In the experiments of learning robot behaviors, the robot was expected to use its gripper to pick up the object in the environment, and then put the
object down outside the working area. For the task here, the input layer includes seven units: six units for the infrared sensors mounted in the front side of the robot, and one unit for the sensor equipped in the inner side of the gripper (to detect the object). The output layer includes four units: two for the left and right motors, and the other two for the packed gripper commands “pick-up” and “release”. A supervised learning algorithm was employed to successfully derive a neural network controller that satisfies the collected data of sensor-motor mapping by minimizing the accumulated motor error. Figure 1 presents the typical behavior produced by the robot after learning.
Figure 1: The robot behavior.
A set of experiments has also been done to verify our approach of emotion-based control. In the experiments, the simple types of emotions are modeled, including “happy”, “angry”, “fear”, “bored”,“shock”,and “sad”.Also threevariables, “hungry”,“tired”,and “familiar”are defined to indicate the internal body states of the robot. As mentioned above, the user is allowed to define event procedures and the relevant weight parameters for the above emotions and body states, to describe how the quantities of different emotions vary over time for their own robots.
With the defined emotion model, a robot can act autonomously. At each time step, the internal emotions and states of the robot change and the newly obtained values are used as the input of an emotion-based coordinator to select behavior controller at that moment. To train a pet robot behaves as expected, a user can build a behavior coordinator to map the combinations of emotions and body states into appropriate behavior controllers. A feedforward neural network is trained from examples to work as a behavior coordinator. Here, the abovementioned variables
(emotions and body states) are arranged as the input of the network, and the output is used to determine which behavior to perform at a certain time. Currently, ten basic behaviors are built. In the training phase, the user is allowed to give a set of training examples in which each example specifies the behavior the robot is expected to perform when the set of emotions and states reaches the values he has assigned. The back-propagation algorithm is then employed to learn a mapping strategy with best approximation from the examples. If the model has been derived successfully but the behavior of the robot has not yet satisfied the owner’s expectation, he can correct the robot behavior for any specific time step by editing the training set through the interface. Then the modified outputs can be used as new training examples to re-train the network to derive a new strategy of behavior arbitration. In this way, the user can easily and conveniently design the characteristics of his personal robot.
To investigate the effect of the number of behaviors to be selected, five sets of experiments have been performed in which different numbers of behaviors were used in the training sets. After each training phase, the pet with the obtained behavior arbitrator was tested for a period of time and its behavior choices were examined and recorded. Figure 2 shows the results. It indicates that when the number of behaviors increases the training task becomes more difficult to achieve.
Figure 2: The percentage of the robot performed correctly
during the test periods.
4. Conclusion
In this work, we have described the importance of developing toy-type pet robots as an intelligent robot application. To realize the development of pet robot, a user-centered interactive framework has been constructed with which the user can conveniently configure and re-configure his personal pet robot according to his preferences. Our framework mainly
0 20 40 60 80 100 1 2 3 4 5 6 7 8 9 10 11 num. of behavior co rr ec tn es s (% )
investigates the issues of distributed robot control and robot emotion, in which the control system for a pet robot is composed of a set of feedforward network networks. To automate the design of control systems, an evolutionary system has been developed for creating new behavior primitives. In addition, an emotion system has been developed in which different emotions and internal states have been modeled and used to derive a behavior arbitrator. The behavior arbitrator is a neural network and the user is allowed to define training examples to infer a personal arbitrator for his robot. To evaluate our framework, we have used it to build robot controllers to achieve various tasks successfully. Reference
[1] Fujita, M. (2004) ‘On activating human communications with pet-type robot AIBO’,
Proceedings of the IEEE, Vol. 92, pp.1804-1813.
[2] Perzarcowski, D., Schultz, A., Adams, W., Marsh, E. and Bugajska, M. (2001) ‘Building a multimodal human-robot interface’, IEEE Intelligent Systems, Vol. 16, pp.16-21.
[3] Fellous, J.-M., Arbib, M. Eds. (2005) Who Needs
Emotions? The Brain Meets the Robot. Oxford
University Press.
[4] Damasio, A. R. (1994) Descarte’sError:Emotion,
Reason, and Human Brain. NY: Grosset/Putnam
Press.
[5] Arkin, R. C. (1998) Behavior-Based Robotics. MA: MIT Press.
5
