Masatoshi Tokita

Skill based control by using fuzzy neural network for hierarchical intelligent control

A novel architecture of an intelligent control system for robotic manipulators is presented. The system is an integrated approach of neuromorphic and symbolic control of a robotic manipulator, including an applied neural network for the servo control, a knowledge-based approximation, and a fuzzy neural network (FNN) for skill-based control. The neural network in the servo control level is the numerical manipulation, while the knowledge-based part is the symbolic manipulation. In neuromorphic control, the neural network compensates for the nonlinearity of the system and the uncertainty in the environment. The knowledge-based part develops the control strategy symbolically for the servo level. The FNN is used between the servo control level and the knowledge-based part to link numerals to symbols and express human skills through learning. This system is analogous to the human cerebral control structure combined with reflex action.<<ETX>>

Neuromorphic control: adaptation and learning

A structure for a neural network-based robotic motion controller is presented. Simulations of both position and force servos are carried out, and the approach is shown to be useful for a nonlinear system in an uncertain environment. The neural network comprises a four-layer network, including input/output layers and two hidden layers. Time delay elements are included in the first hidden layer, so that the neural network can learn dynamics of the system. The authors also implement a new learning method based on fuzzy logic, which is useful to accelerate learning and improve convergence. >

Neural network application for robotic motion control-adaptation and learning

The authors consider neural network applications to robotic motion control in which the controller is used for the position and force control of robotic manipulators. The proposed neural servo controller is based on a neural network consisting of two hidden layers and input/output layers. The controller can adjust the neural network output to the robot in the forward manner to cooperate with the feedback loop, depending on unknown characteristics of handling objects. In particular, the proposed neural network has delay elements in itself, so that it can learn the dynamics of the system. Simulations are carried out for the case of one- and two-dimensional robotic manipulators. The performance of the proposed neural servo controller is shown in terms of its frequency response, and the robustness against impulsive noises is also shown. The authors propose a fuzzy turbo to avoid stagnation, so that the neural network can learn the dynamical system quickly

New strategy for hierarchical intelligent control of robotic manipulator-hybrid neuromorphic and symbolic control

The authors present a novel scheme for intelligent control which is an integrated approach of the neuromorphic and symbolic control of a robotic manipulator, including an applied neural network (NN) for servo manipulation, while the knowledge-based part is symbolic manipulation. This control system is analogous to the human cerebral control structure. Experiments and simulations demonstrated that the NN can connect the numerical real world with the symbolic artificial intelligence world.<<ETX>>

Hybrid symbolic and neuromorphic control for hierarchical intelligent control

The authors present a scheme for intelligent control of robotic manipulators. This is a hybrid system of neuromorphic control and symbolic control of a robotic manipulator, including a neural network for the servo control and a knowledge-based approximation. The neural network in the servo control level is for numerical manipulation, while the knowledge-based approximation is for symbolic manipulation. In neuromorphic control, the neural network compensates for the nonlinearity of the system and the uncertainty in the environment. The knowledge base develops the control strategy symbolically for the servo level. This is an analogous control system to the human cerebral control structure.<<ETX>>

Adaptation and learning for hierarchical intelligent control

The authors discuss a novel strategy for hierarchical intelligent control. They propose this strategy for a neural-network-based controller to be generalized with the higher level control based on artificial intelligence and to acquire knowledge heuristically. This system comprises two levels: a learning level and an adaptation level. The neural networks are used for both levels. The learning level has a hierarchical structure for recognition and planning, and is used for the strategy of robotic manipulation in conjunction with the knowledge base in order to expand the adaptive range. The recent information from the adaptation level updates the learning level through a long-term learning process. On the other hand, the adaptation is used for the adjustment of the control law to the current status of the dynamic process.<<ETX>>

Force control of a robotic manipulator by application of a neural network

In this paper, a force control method for robotic manipulators which utilize a neural network model is proposed with consideration of the dynamics of objects. The proposed system consists of a standard PID controller and a multilayered neural network model, which optimizes a set of controllers parameters via a process of learning. The neural network model has not yet been applied to force control problems, but the proposed method is shown to be applicable to force/compliance control problems. The stability of this system and a wider applicability are verified by simulation studies.

Neuromorphic sensing and control-impact control of robotic manipulator

A new approach for impact control of a robotic manipulator by a neural network (NN)-based controller is presented. Collisions are very quick phenomena and have strong nonlinearity. Therefore, it is difficult to sense collisions and to control a robotic manipulator undergoing collisions. The proposed approach has robustness against the impact force. It also has capabilities for sensing and perception of collisions by using a proximity sensor. In this case, the NN-based controller can use a nonlinear feedback in approaching the object to avoid impact. The NN-based control acquires these capabilities of its own accord.<<ETX>>

Neuromorphic sensing and control for robotic manipulator-position, force and impact control

Summary form only given. The authors discuss a neural-network (NN)-based robotic motion control system, the neuromorphic control system, employing the neural servo controller (NSC). The NSC uses the NN including time delay elements in the hidden layer so as to learn the dynamics without information about accelerations. Simulations of both position and force control are carried out with consideration of nonlinear collision phenomena. The proposed approach is shown to be particularly valuable in sensing and recognition of collision phenomena employing proximity sensors, and in control of nonlinear systems with unknown parameters. A Hertz-type model with energy loss parameter was adopted to express the impact force. It was also shown that, since the NSC can sense conditions, the NSC can achieve nonlinear feedback both in position control mode and in force control mode to avoid the influence of collisions.<<ETX>>

Hierarchical hybrid neuromorphic control system for robotic manipulator

Summary form only given, as follows. The authors discuss a hierarchical hybrid neuromorphic control system for intelligent control of robotic manipulators and for acquiring knowledge and learning. This robotic motion control system comprises two levels: a learning level and an adaptation level. Neural networks are employed for both levels. The learning level has a hierarchical structure and is used for the strategic planning of the robotic manipulator in conjunction with the neural knowledge database (NKDB) in order to enlarge the adaptation range. The proposed NKDB can infer an unknown fact from prior knowledge for strategic planning and is updated by the recent information from the adaptation level, the neural servo controller, through a long-term learning process. On the other hand, the adaptation is used for the adjustment of the control law to the current status of the dynamic process. The initial states of the adaptation level are given by the NKDB. Experiments and simulations show that, if the NKDB recognizes objects correctly, the NSC can adapt easily, but if not the NSC tries to adapt objects and gain the force sensing information so as to acquire new knowledge for the NKDB.<<ETX>>