Feedback Error Learning Research Articles

SIMBICON

Physics-based simulation and control of biped locomotion is difficult because bipeds are unstable, underactuated, high-dimensional dynamical systems. We develop a simple control strategy that can be used to generate a large variety of gaits and styles in real-time, including walking in all directions (forwards, backwards, sideways, turning), running, skipping, and hopping. Controllers can be authored using a small number of parameters, or their construction can be informed by motion capture data. The controllers are applied to 2D and 3D physically-simulated character models. Their robustness is demonstrated with respect to pushes in all directions, unexpected steps and slopes, and unexpected variations in kinematic and dynamic parameters. Direct transitions between controllers are demonstrated as well as parameterized control of changes in direction and speed. Feedback-error learning is applied to learn predictive torque models, which allows for the low-gain control that typifies many natural motions as well as producing smoother simulated motion.

SIMBICON

Physics-based simulation and control of biped locomotion is difficult because bipeds are unstable, underactuated, high-dimensional dynamical systems. We develop a simple control strategy that can be used to generate a large variety of gaits and styles in real-time, including walking in all directions (forwards, backwards, sideways, turning), running, skipping, and hopping. Controllers can be authored using a small number of parameters, or their construction can be informed by motion capture data. The controllers are applied to 2D and 3D physically-simulated character models. Their robustness is demonstrated with respect to pushes in all directions, unexpected steps and slopes, and unexpected variations in kinematic and dynamic parameters. Direct transitions between controllers are demonstrated as well as parameterized control of changes in direction and speed. Feedback-error learning is applied to learn predictive torque models, which allows for the low-gain control that typifies many natural motions as well as producing smoother simulated motion.

Seasonal variation of neuro-controller training period observed in the control task of inverted pendulum

In general, computer simulations are a useful means to confirm the effectiveness of a control scheme, but the actual system is not so easy to handle. In order to overcome such underlying realistic problems, adaptability of neural networks is focused on by many researchers. In this study, an on-line training model for self-turning control is constructed based on the famous feedback error learning model, and its effectiveness is investigated through a task for controlling an inverted pendulum. As a result of several experiments, repeatable strange phenomena, such as seasonal variation of training period and reversalization of control command, are observed.

Mechanisms of cerebellar gait ataxia

The cerebellum is important for movement control and plays a critical role in balance and locomotion. As such, one of the most characteristic and sensitive signs of cerebellar damage is gait ataxia. How the cerebellum normally contributes to locomotor behavior is unknown, though recent work suggests that it helps generate appropriate patterns of limb movements, dynamically regulate upright posture and balance, and adjust the feedforward control of locomotor output through error-feedback learning. The purpose of this review is to examine mechanisms of cerebellar control of locomotion, emphasizing studies of humans and other animals. Implications for rehabilitation are also considered.

Role of cerebellum in learning postural tasks

For a long time, the cerebellum has been known to be a structure related to posture and equilibrium control. According to the anatomic structure of inputs and internal structure of the cerebellum, its role in learning was theoretically reasoned and experimentally proved. The hypothesis of an inverse internal model based on feedback-error learning mechanism combines feedforward control by the cerebellum and feedback control by the cerebral motor cortex. The cerebellar cortex is suggested to acquire internal models of the body and objects in the external world. During learning of a new tool the motor cortex receives feedback from the realized movement while the cerebellum produces only feedforward command. To realize a desired movement without feedback of the realized movement, the cerebellum needs to form an inverse model of the hand/arm system. This suggestion was supported by FMRi data. The role of cerebellum in learning new postural tasks mainly concerns reorganization of natural synergies. A learned postural pattern in dogs has been shown to be disturbed after lesions of the cerebral motor cortex or cerebellar nuclei. In humans, learning voluntary control of center of pressure position is greatly disturbed after cerebellar lesions. However, motor cortex and basal ganglia are also involved in the feedback learning postural tasks.

Discrete-Time Feedback Error Learning and Nonlinear Adaptive Controller

In this study, the technique of Discrete-Time Feedback Error Learning (DTFEL) is investigated from the viewpoint of adaptive control. First, the relationship between DTFEL and nonlinear discrete-time adaptive control with adaptive feedback linearization is discussed. It shows that DTFEL can be interpreted as a form of nonlinear adaptive control. Second, Lyapunov analysis of the controlled system is presented. It suggests that the condition of strictly positive realness (s.p.r.) associated with the tracking error dynamics is a sufficient condition for asymptotic stability of the closed-loop dynamics. Specifically, for a class of second order SISO systems, it is shown that the control problems is reduced to select the feedback gain that satisfies the s.p.r. Finally, numerical simulations is presented to illustrate the stability properties of DTFEL obtained from mathematical analysis.

Feedback Error Learning for Writing One Stroke Characters by Two-Link Manipulator

In this paper, we consider a version of problem how to teach robots to write characters in actual environment. In particular, one must design a feedforward controller for two-link manipulators to improve the tracking performance in the face of limited knowledge of the surroundings. We employ an adaptive scheme, called MIMO-FEL (Multi-Input Multi-Output Feedback Error Learning) to achieve our objective. The effectiveness of proposed method is demonstrated with an experiment.

Closed-loop Identification Based on MIMO Feedback Error Learning

Motivated by the recent development of Feedback Error Learning (FEL), this paper proposes a method for closed-loop identification of a Multi Input Multi Output (MIMO) plant. Given a control system roughly designed, a feedforward controller is constructed by learning to achieve desirable responses. Then the learned feedforward controller gives a model of the plant, which will be effective for re-designing the control system to improve the performance. The effectiveness of the method is verified through numerical simulation.

Generalization of Feedback Error Learning (FEL) to MIMO Systems

Generalization of Feedback Error Learning (FEL) to MIMO Systems

Neural Correlates of Internal-Model Loading

Skilled object manipulation requires knowledge, or internal models, of object dynamics relating applied force to motion , and our ability to handle myriad objects indicates that the brain maintains multiple models . Recent behavioral studies have shown that once learned, an internal model of an object with novel dynamics can be rapidly recruited and derecruited as the object is grasped and released . We used event-related fMRI to investigate neural activity linked to grasping an object with recently learned dynamics in preparation for moving it after a delay. Subjects also performed two control tasks in which they either moved without the object in hand or applied isometric forces to the object. In all trials, subjects received a cue indicating which task to perform in response to a go signal delivered 5-10 s later. We examined BOLD responses during the interval between the cue and go and assessed the conjunction of the two contrasts formed by comparing the primary task to each control. The analysis revealed significant activity in the ipsilateral cerebellum and the contralateral and supplementary motor areas. We propose that these regions are involved in internal-model recruitment in preparation for movement execution.

A new control scheme for nonlinear systems with disturbances

A new learning control scheme, based on a nonlinear disturbance observer (NDO) coupled with a sliding-mode fuzzy neural network (SFNN) with a feedback-error-learning (FEL) strategy, is proposed for a class of time-varying nonlinear systems with unknown disturbances. The proposed controller, referred to as NDOFEL, involves two steps for obtaining an estimate of the time-varying lumped disturbance d(t) for improving the precision of the tracking control. The NDO is initially applied to estimate d(t), but an observer error does not converge to zero since d/spl dot/(t)/spl ne/0. The SFNN is then presented to estimate the observer error such that the output of systems follows a desired trajectory. The proposed NDOFEL has stable on-line learning ability, maintains high control performance in the presence of disturbance, and guarantees the stability of closed-loop systems on the basis of the Lyapunov theorem. The effectiveness and robustness of the proposed NDOFEL is demonstrated through simulation results obtained for the tracking control during wing rock phenomena. The results suggest that the proposed controller can significantly enhance the tracking performance of aircraft.

MIMO Feedback Error Learning:Algorithm and its Stability

MIMO Feedback Error Learning:Algorithm and its Stability

Nonlinear Adaptive Neural Controller for Unstable Aircraft

A model reference indirect adaptive neural control scheme that uses both off-line and online learning strategies is proposed for an,unstable nonlinear aircraft controller design. The bounded-input-bounded-output stability requirement for the controller design is circumvented using an off-line, finite interval of time training scheme. The aircraft model is first identified using a neural network with linear filter (also known as time-delayed neural network) with the available input-output data for a finite time interval. The finite time interval is selected such that this time interval is less than the critical time interval for the aircraft from its stability point of view (similar to the time to double). A procedure to select this critical time interval is also presented. For a given reference model and the identified model, the controller neural-network weights are adapted off-line for the same time interval. The off-line trained neural controller ensures the stability and provides the necessary tracking performance for the unstable aircraft. If there is a change in the aircraft dynamics or characteristics, the trained neural identifier and controller are also adapted online. The theoretical results are validated using the simulation studies based on a locally nonlinear longitudinal high-performance fighter aircraft similar to the F-16. The neural controller design proposed is also compared with the feedback error learning neural control strategy in terms of the tracking ability and control efforts for various level flight conditions and fault conditions such as modeling uncertainties and partial control surface loss. We also present the robustness of the aircraft under extreme wind and noise conditions.

Stability of feedback error learning method with time delay

Feedback error learning method was recently proposed by Kawato et al. [A heirarchical neural-network model for control and learning of voluntary movements, Biol. Cybernet. 57 (1987) 169–185.] as a possible architecture of brain motor control which is supported by experimental results in neurophysiology. In this paper, we analyze it as a two-degree-of-freedom adaptive control for general time invariant linear plant with adaptive controller in the feedforward path. A time delay is allowed in the feedback loop as in the neuronal pathways of motor control. We derive stability condition of the feedback error learning method based on the strict positiveness of the closed loop system. The control performance of the feedback error learning method as a design strategy of adaptive control has been demonstrated by simulation results.

Application of a multi-DOF ultrasonic servomotor in an auditory tele-existence robot

A multi-degree-of-freedom (DOF) ultrasonic motor can rotate in three DOFs and does not generate noise. In addition, with an appropriate preloading mechanism, it can generate high torque for its size. The multi-DOF ultrasonic motor is, therefore, anticipated for use as a servomotor in the next generation of robots. However, for several reasons, there have been few applications of multi-DOF ultrasonic motors. One reason is the difficulty in designing a proper preloading mechanism for the motor and the limitation of the size of the stators. Another is the difficulty in developing a control algorithm, due to the motor's complex and changing dynamical characteristics, and the serious jaggy motion caused by its very quick response. This paper proposes a preloading mechanism and control algorithm for a multi-DOF ultrasonic motor, considering the motor's application to an actual auditory tele-existence robot, TeleHead. TeleHead is an elaborate dummy head robot that has a 3-DOF neck mechanism. The proposed methods achieve smooth and fast multi-DOF rotating motion of the dummy head with little time delay. In the preloading method, tensioned springs generate high preloading force and make up for the lack of torque by compensating for the resistance torque generated by the inclining motion of the dummy head. In the control algorithm, high-DOF motion is managed by high-frequency switching of the rotating axis, and smooth and quick trajectory tracking motion is achieved by introducing feed-forward control using an inverse model, with multiresolution acquired by feedback-error learning. Experimental results verify the high performance of these methods.

Joint angle control by FES using a feedback error learning controller

The feedback error learning (FEL) scheme was studied for a functional electrical stimulation (FES) controller. This FEL controller was a hybrid regulator with a feedforward and a feedback controller. The feedforward controller learned the inverse dynamics of a controlled object from feedback controller outputs while control. A four-layered neural network and the proportional-integral-derivative (PID) controller were used for each controller. The palmar/dorsi-flexion angle of the wrist was controlled in both computer simulation and FES experiments. Some controller parameters, such as the learning speed coefficient and the number of neurons, were determined in simulation using an artificial forward model of the wrist. The forward model was prepared by using a neural network that can imitate responses of subject's wrist to electrical stimulation. Then, six able-bodied subjects' wrist was controlled with the FEL controller by delivering stimuli to one antagonistic muscle pair. Results showed that the FEL controller functioned as expected and performed better than the conventional PID controller adjusted by the Chien, Hrones and Reswick method for a fast movement with the cycle period of 2 s, resulting in decrease of the average tracking error and shortened delay in the response. Furthermore, learning iteration was shortened if the feedforward controller had been trained in advance with the artificial forward model.

A fault-tolerant neural aided controller for aircraft auto-landing

This paper presents a neural-aided controller that enhances the fault tolerant capabilities of a high performance fighter aircraft during the landing phase when subjected to severe winds and failures such as stuck control surfaces. The controller architecture uses a neural controller aiding an existing conventional controller. The neural controller uses a feedback error learning mechanism and employs a dynamic Radial Basis Function neural network called Extended Minimal Resource Allocating Network (EMRAN), which uses only on-line learning and does not need a priori training. The conventional controller is designed using a classical design approach to achieve the desired autonomous landing profile with tight touchdown dispersions called herein as the pillbox. This design is carried out for no failure conditions but with the aircraft being subjected to winds. The failure scenarios considered in this study are: (i) Single faults of either aileron or elevator stuck at certain deflections, and (ii) double fault cases where both the aileron and elevator are stuck at different deflections. Simulation studies indicate that the designed conventional controller has only a limited failure handling ability. However, neural controller augmentation considerably improves the ability to handle large faults and meet the strict touchdown dispersion requirements, thus enlarging the fault-tolerance envelope.

PD 기반의 퍼지제어기로 제어된 로봇의 새로운 신경회로망 보상 제어 기술

In this paper, a novel neural network compensation technique for PD like fuzzy controlled robot manipulators is presented. A standard PD-like fuzzy controller is designed and used as a main controller for controlling robot manipulators. A neural network controller is added to the reference trajectories to modify input error space so that the system is robust to any change in system parameter variations. It forms a neural-fuzzy control structure and used to compensate for nonlinear effects. The ultimate goal is same as that of the neuro-fuzzy control structure, but this proposed technique modifies the input error not the fuzzy rules. The proposed scheme is tested to control the position of the 3 degrees-of-freedom rotary robot manipulator. Performances are compared with that of other neural network control structure known as the feedback error learning structure that compensates at the control input level.

516 テニス時の筋力特性を考慮した人間系とラケットとボールの連成挙動解析

The analysis of coupled vibration of the arm, the tennis racket and the ball is important to understand the performance of the racket and to prevent of sports diseases when people play tennis. The arm is modeled as musculo-skeletal model taking activity level into account, and the ball and the racket are modeled by masses, rigid links, respectively, which are connected by springs and dampers. By coupling the dynamics of the ball, the racket, and the arm, the coupled equation of motion is derived. The feedback error learning system is introduced to obtain the value of the activity level of muscle and the holding moment between hand and racket. Finally the parameter study is performed, we show dynamic interaction among the racket, the ball and the arm.

1405 筋力特性を考慮したテニスラケット・ボール・人間系の運動と振動(GS-10 スポーツ工学,研究発表講演)

The analysis of coupled vibration of the arm, the tennis racket and the ball is important to understand the performance of the racket and to prevent of sports diseases when people play tennis. The arm is modeled as musculo-skeletal model includiing activity level, and the ball and the racket are modeled by masses, rigid links, respectively, which are connected by springs and dampers. By coupling the dynamics of the ball, the racket, and the arm, the coupled equation of motion is derived. The feedback error learning system is introduced to obtain the activity level of muscle. And we calculate the holding moment between hand and racket from feedback error learning system. Finally this proposed analysis method is compared with the former analysis and experimental results.