A power assisting device for lower back flexion and extension, when carrying a heavy load is presented in this paper. When people lift up something heavy, they use both hands. So the development of a controller without using hands is necessary. We attempt to control the device by a voluntary human motion. First, lifting up and putting down motions are analyzed using a human model. Results show that torques in a hip joint and a knee joint are characteristic for control the device. It means that EMG signals in front and back thigh muscles can become feature values because the muscles located there connect the hip joint and the knee joint. After that, a controller of the device using the EMG signals is developed. Finally, artificial neural networks (ANN) are introduced as a solution of the problem of individual differences.

In this paper, we propose a power assist device for lower back flexion and extension, when carrying a heavy load. To see the effect of the device, we model a human body and analyze a compression force in his lower back, as well as the evaluation of a supported force at a hand position. A prototype of the device is manufactured and a controller of the device is developed, which can follow a voluntary human motion assisting his strength.

Reconfigurable robots have functional flexibility and physical flexibility. A tripod robot is one of the robots having SMA actuators and it is characterized by its symmetrical structure and reconfigurability. When realizing movements of autonomous robots, the model of both the robot and the environment am required for the controller. In this paper, the model for control the tripod robot is generated as a logic circuit by Net-list evolution. The Net-list chromosome is one of the circuit models that abstracted a logic circuit structurally. The generation of the adaptive behavior of the tripod robot and the control model are discussed through the experiment of using a simulator.

In observing behavior of multiple autonomous robots, a microscopic observation expressed by dynamic equations is usually used. However, it is very difficult to estimate behavior of robots or mutual interactions among them in real time. Furthermore, it is hard to realize the observed system by taking consideration in all the factors of the system. On the other hand, a macroscopic observation defined by state equations is efficient for recognizing behavior of multiple robots. In this study, a quantitative observation approach is proposed on behavior of multiple robots. This approach introduces macroscopic state quantities in thermodynamics into expression of the multiple robots system. The advantage of this approach is that observation on the behavior of autonomous robots in real world can be mapped to characteristic quantities in another conceptual state space. At first, the state quantities of multiple robots system are defined and their physical meaning is discussed regarding it as quantitative observation of multiple robots. The macroscopic state quantities in thermodynamics, such as temperature, pressure and entropy, are introduced into mobile robots system. Each mobile robot is regarded as a particle in term of thermodynamic systems. Experiments show that the states of robots system can be classified from a viewpoint of the macroscopic state quantities in thermodynamics. This verifies that the macroscopic quantitative observation is efficient and applicable to controlling multiple robots system.

In order to achieve an autonomous system which can adaptively behave through learning in the real world, we have constructed a distributed autonomous swimming robot that consists of mechanically linked multi-agent and adopts adaptive oscillator method that was developed as a general decision making for distributed autonomous systems. One of the aims of using this system is to verify whether the robot could complete a target approaching including obstacle avoidance. For this purpose, we introduce a modified Q-learning in which plural Q-tables are used alternately according to dead-lock situations. By using this system, as a result, the robot acquires a stable target approaching and obstacle avoiding behavior.

The purpose of the article is the development of an interface which closely adapts to the individual. By quantifying the frustration as the human manipulates machines from a biomedical signal and making it into teaching signals of machine learning (ML), we aim at the development of a system in which the machine adapts to the human. The authors extract the characteristic vector of whether the examinee is in discomfort or not from electroencephalograms (EEG) and electromyograms (EMG). An artificial neural network (ANN) is employed to extract the vector. For the machine learning, reinforcement learning is used and the rewards are an extracted signal from physiological signals. As a basic experiment for extracting discomfort and comfort from the physiological signals, the EEG measurement experiment was carried out under an unpleasant sound environment for 20 examinees. The input signals to ANN for the characteristic vector extraction was examined. By affixing piezoelectric films on the eyebrow, the movement of the eyebrow was measured. Finally, the results of measuring EEG and EMG simultaneously under the situation in which frustration accumulated for the examinee are shown. We use reinforcement learning (RL) to control the behavior of the Khepera robot.

Proposes the construction of a "strong" autonomous mobile robot, which can acquire environment oriented behavior through learning, as a distributed autonomous system. It is thought that such a system has many advantages over other systems in terms of adaptability to the environment and so on. However, the potential of this type of system has yet to be demonstrated in experiments under real-world conditions. We conducted an experiment to determine whether a distributed autonomous swimming robot could acquire target-approaching behavior on a water surface which was set as the robot's work space. As a result, from a fairly simple coding, the robot acquired the reproducible target-approaching behavior using only local learning even in cases where a partial fault occurred, and the acquired actions also enabled the robot to approach the target in an environment with a narrow gate.

We discuss the necessity of a learning mechanism for an EMG prosthetic hand controller, and the real-time learning method is proposed and designed. This method divides the controller into three units. The analysis unit extracts useful informations for discriminating motions from the EMG. The adaptation unit learns the relation between EMG and control command and adapts operator's characteristics. The trainer unit makes the adaptation unit learn in real-time. Experiments show that the proposed controller discriminates ten forearm motions, which contain four wrist motions and six hand motions, and learns within 4/spl sim/25 minutes. The average of the discriminating rate is 91.5%.

Presents an evolutionary method for robotics. Reactive architecture leads robots to move in the real world adaptively according to a quick response. However, designing adequate complete rules for complex behavior is practically impossible, because interaction among components is unlikely to be completely known. To approach this difficult design problem in the arbitration of the interaction organizing a system, a new approach "gate growth", which is a grammar encoding network construction method, is presented. The methodology consists of two components: a cellular automaton (CA) as a connection algorithm and an NAND array as a device model. Through genetic algorithm operations, rules forming the CA are evolved to achieve given tasks. In application to a wall following problem with Khepera, the possibility of building task achieving behavior out of the gate growth is shown.

This study presents an autonomous agent which is able to behave appropriately in the complex environment with unsophisticated sensors. To satisfy this aim, an internal cognitive space, called a metamorphic internal image space, is introduced. The metamorphic internal image space is a space that integrates the sensory inputs and the temporal change of the space, so that the agent is able to behave by relying only on the image. An agent with the metamorphic internal image is considered to have following two major characteristics: 1) it can learn the chain of actions considered as a behavior; and 2) it can move with/without the change of sensory input pattern. A learning mechanism based on a multi-agent system is introduced for the realization of the metamorphic internal image. The behavior of the multi-agent system can be regarded as a complex system. Computational experiments show the emergent behavior of the mobile agent with metamorphic internal image based on the complex learning system.

The environment in the real world has an almost infinite number of events. From a Cartesian point of view, these events consist of more detailed factors. Those factors often correspond to the physical properties of the environment and this renders symbolic AI powerless. However, affordance, i.e. the availability of behaviours with respect to the environment, uses functional properties of which it is possible to be directly aware. To make the agent more adaptive, the useful information is obtained by the agent itself. This paper describes two perception mechanisms that have possibilities to allow the agents to perceive the affordance through learning. The mechanism is aimed for the perception of the environment and the self-perception, respectively. The computer simulations and some preliminary experiment by the real agent, a vehicle with vision system, show the possibilities of the mechanisms.

It is well known that subsumption architecture (SSA) gives a great impact on the robot design methodology. However there are still some shortfalls on SSA due to the lack of learning ability. An SSA with learning ability is expected to produce a more powerful SSA, in terms of flexibility, robustness, etc. Here a new extension of SSA is proposed, which is a kind of distributed SSA with behaviour networks as the learning function. This specialized distributed SSA behaves like multi-autonomous agents. The proposed method can be partitioned into two layers: an upper layer and a lower layer. The upper layer performs the learning of mimic impression and in turn commands the SSA agents in the lower layers. Through the real world experiment the effectiveness and adaptability of the proposed architecture are shown.

Attempts to realize an autonomous solving mechanism for a task planning problem by a machine learning system. As a robot task planning problem, the block stacking problem is treated. It is well known as one of the most difficult problems. In particular the difficulties of this problem increase in a multiagent environment where multiple robots exist in a problem domain and they cooperate or negotiate to achieve given tasks effectively. To realize an autonomous planning mechanism, a classifier system is applied to this problem. In this approach, a classifier corresponds to a production rule that instructs the next operation when its conditional part is fully matched for a current state. Each robot has an individual classifier system as the robot task planner.

This paper proposes potential field techniques for control of a link mechanism based on a decentralized management and an approach to describing a mathematical model of autonomous machines. To control the group of machines, there are two categories of method, one was based on central control and the other on parallel distributed control. A popular method for control machines is the so-called potential method, which has been applied to navigation for automatically guided vehicles in a stable working space. Since the potential function of the working space must be determined at every instant for an unstable working space, an efficiency of the potential method becomes low. To modify the potential function local, an approach called the vibrating potential method (VPM) is introduced. With this method, there are many cases where high-performance information processing can be attained without supervision. This paper proposes a new control method based on VPM and discusses its application to the grasping problem of free form objects.

A new approach of how to solve the 3D packing problem automatically is proposed. This problem is well known as a complete combinatorial problem. The authors attempt to realize a mechanism in which the 3D packing rule is automatically tuned, and an optimum packing solution is obtained by applying genetic algorithms which mimic the process of a natural evolution system. The 3D packing strategy is controlled by two evaluation functions which dominate the selection of a next allocation position and a box. To find the near-optimal strategy, the weighted coefficients of the evaluation functions are tuned by applying the genetic operators such as reproduction, crossover and mutation. To use the obtained tuned results as accumulated successful strategies, a 3D packing rule-base is constructed. The rules in this rule-base are composed of a 'conditional part', which expresses the features of the given problem, and a 'procedural part', which gives the packing strategy.>

Describes a genetic algorithm (GA) based approach to realize expert knowledge acquisition and refinement in a real world problem: parameter value determination for metal shaft grinding where several trade-offs exist among operational conditions to obtain expected surface finish. Causal relations are encoded to strings in (condition): (action) form and implemented as a rule based system. Genetic operators are applied to: (1) discretization of continuous parameter values; and (2) refinement of rules.>