Shigeto Aramaki

Development of autonomous mobile humanoid robot

The authors built a humanoid robot which behaves the same as the human in the space concerned. The mechanism and functions, such as vision sensor, acoustic sensor and loudspeaker, are mounted on this robot in order that the multi modal bidirectional communication between the robot and the human is possible. The authors adopted parallel processing by multi CPUs. The computers in the robot body are mutually connected by an Ethernet LAN. This LAN consists of the cerebrum system LAN and the motion system LAN in which a robot can quickly perform conditioned reflex actions as well as a human does. By using a wireless LAN, the LAN in a robot is connected to the outside LAN connecting the computers for software development and system support. In this paper, the outline of the developed humanoid robot is described.

Online Cooperative Behavior Planning Using a Tree Search Method in the RoboCup Soccer Simulation

In this paper, we propose a tree search approach to generate and evaluate cooprative behavior online in multiagent systems. It was difficult to apply a tree search methodology to tasks that the state-action space is continuous and requires realtimeness. However, it has become possible to apply such an approach since the computational resources became more powerful today. We applied a tree search method to the Robo Cup soccer 2D simulation and analyzed its effect by evaluating the team performance.

The knowledge representation and programming for robotic assembly task

Presents a knowledge representation and programming method for robotic assembly tasks. The methodology is explained through robotic assembly tasks with position and force control. The state space of assembly objects is abstractly defined using a qualitative landmark which consists of the state of an object and a robot. The current state and sensor input determines the next robot action, and the state of a robot is made to lead to the goal state. These concepts are formulated by using a finite state automaton. The action sequence is defined based on the control hierarchy structure which consists of four levels (control primitive level, control skill level, skill level and task level). The complicated assembly task can be represented by product and concatenation of control primitives in a lower level and various tasks can be defined by adding new control primitives. In the paper, the concept of a CRS (constraint reduction system) is introduced for making a robot task program. The CRS is realized by extended Prolog. All the action and primitives are considered as a process and can be uniformly described by the reduction rule of the CRS.

Learning Evaluation Function for Decision Making of Soccer Agents Using Learning to Rank

In the simulated soccer domain, the evaluation of actions that can be performed by soccer agents is important to design the team behavior. It is not easy to design the evaluation function that reflects the intention of human developers. Usually, manual adjustment of the evaluation function requires much cost. We propose a method to apply Learning to Rank algorithm to the evaluation function of the decision making of soccer agents. The experimental results show Learning to Rank would be promising to using the instruction of human trainer.

A Robot Programming Based on Frame Representation of Knowledge

We have developed the method of a robot programming by using frame like knowledge base in order to compose the multi-modal human-robot interface. The concept of object oriented programming, the case grammar, and the Conceptual Dependency theory (CD theory) is introduced into this knowledge base in order that the total system can be easily and naturally composed. In this system, a robot can interact with humans in the voice and the finger pointing. Furthermore, termination condition of a robot task can be easily realized by using CD theory, and the robustness of robot tasks is improved. We carried out the work experiment by using actual humanoid robot and the effectiveness was confirmed.

An Assembly Structure using Functional Element for Product Assembly Sequence Generation

Themanagement of design information is been becoming more and more important, and considerable research is being carried out on the representation of design information of assembly parts. This paper describes the representation of an assembly structure, which includes the introduction of a functional element. By using this functional element, constraint of the assembly structure can be naturally described. The functional element is used to express relations among the parts, and a relation can describe constraints among assembly parts which interfere without touching. The relation includes an assembly method and is capable of describing both constraint and interference. We demonstrate that our representation of the assembly structure reduces the search space of the assembly sequence.

Team formation estimation using cluster analysis and triangulation model

In this paper, we propose a method to estimate unknown team formation from observed positional data. We combined Growing Neural Gas Network and triangulation based function representation model in order to analyze positional data and recompose a team formation. In the experiments, we show the estimation results using the log file recorded by the RoboCup soccer 2D simulator.

The representation method of robotic assembly task with click action

The skill of robotic task has been studied in many research institutes. We have been researching the skill for the assembly of the consumer products. We has developed the original knowledge representation for robotic assembly task with the position and force control. We proposed the representation method of the skill that a human has and the construction method of a knowledge base for robotic assembly task. In this paper, we analyzed the assembly task with the click action. And, we proposed the knowledge representation that checked the differential value of the sensor. As a result, the representation of the assembly task with the click action became possible. In addition, we programmed the proposed representation and confirmed the assembly work in an actual robot with the click action. Then, the effectiveness of the representation was confirmed.

Control program structure of humanoid robot

In this paper, the new control program structure for humanoid robot which consists of multi-CPU, multi-OS and multi-task is described. The method for planning action from multiple sensor information by using the production system is also described. Task program has hierarchical structure like as human being has conscious action and unconscious action. Our design concept is generality of task. We want to make a control program independently on the architecture of a robot and other tasks. We also want to make software architecture simple. Task program is divided into three levels. Tasks must communicate each other. However, if one task can communicate another task at random, the amount of communication increases exponentially. Therefore, we made a rule of task in order to make the amount of communication linearly increase. On the other hand, a robot must determine next action from multiple sensor information. We introduced the reasoning mechanism of production system for determining next action. It becomes easy to determine next action from multiple sensor input by using this method. We could make a robot grasp a block by using this method and simulator, and we confirmed the effectiveness of methods mentioned above.

Fast Pseudo Random Forest Using Discrimination Hyperspace

In recent years, machine learning technique has been applied to various problems. The improvement of computational power enables the processing of large scale data in a practical time and brought the success of machine learning technique. However, the processing speed of current machine learning models still have a potential to be improved. We are trying to improve the processing speed of Random Forest, which is known as a fast and reliable classification model. In this study, we propose a Discrimination HyperSpace called DHS, which realized a pseudo Random Forest. Experiment results show our method runs much faster than original Random Forest without losing classification performance.