Identification of Human Feedforward Control in a Cyber Grasp and Twist Task
Abstract:
When mechanical linkages are replaced by electronic communication and control systems, certain undesirable phenomena can arise that may be difficult to anticipate. In simple cases the lack of dissipativity in the cyber link unmasks instabilities that were present but suppressed in the system with a mechanical (physical) link. In more complex cases, there may be no simple physical equivalent to the system containing the cyber link, and a damping coefficient or dissipative element may not be identifiable. The major goal of this project is to develop appropriate models and accompanying control theory to describe the phenomena that arise in cyber-physical systems that do not have simple correlates in purely physical systems. The “cyber” system of particular interest for us in this work is a system containing a human in the loop. We have been calling these “cyber-physiological” systems, to highlight the role that biomechanics plays in system behavior. But even further, the physiological part of such systems contains control loops implemented in very complex and not-to-be-deciphered neural circuits. We seek appropriate models that capture phenomena observable in cyber-physiological systems, especially those related to the realization of certain “links” in neural circuits, certain links in electronic circuits (and embedded control code), and yet other links in physical components. To reduce the problem to a tractable form, we make certain assumptions based on empirical evidence concerning the timing of feedback control and the mix of feedback and feedforward processes within the human central and peripheral nervous system. We apply system identification techniques and modeling principles guided by control to obtain models which are as simple as possible yet competent to capture empirically observed phenomena. We are currently using a prototypical manual task, the grasp and twist task, which involves a significant amount of coordination that produces interesting observable phenomena under perturbations. In terms of the grasp and twist task, our goals are to capture all the necessary aspects of motion planning, force control, biomechanics, and appropriate mixes of feedforward and feedback control in a simple model. The grasp and twist task is designed using our single-axis rotary haptic device involving a programmable load force. The load force may be programmed to behave like a weight, spring, or damper. We also outfitted the device handle with a force sensor to measure grip force. We have developed a simple model of the human user incorporating a motion source to capture user intent (the neural command) and a spring-damper coupler with mass to capture the mechanics of muscle, tendon, and soft tissues. We use independent system identification experiments to identify the mass, stiffness, and damping parameters of this model. These experiments are independent in that they are performed prior to the grasp and twist task. We have determined, using empirical results and a study of model properties, that it is necessary to use mass, stiffness, and damping parameters that depend on a variable grasp force to capture the observed behaviors using our model of limb mechanics. That is, it is not possible to obtain a motion source trajectory that produces the observed behaviors with parameters that do not depend on grip force. Naturally, the grip force does change during the grasp and twist task and is available through our force sensor. From the data in cyber grasp and twist tasks, we found the fine coordination between load torques and grip forces is preserved. Human users can generate appropriate grip force to achieve the desired movement despite different load torques. This suggests that human users build an internal model of the external environment as well as the limb dynamics. From the experiment data of the grasp and twist task with unexpected load changes, we found users will prepare for the current trial based on the experience of the previous trial. Hence if the load torque changes, users will erroneously prepare for the previous load torque and hence the grip force will be excessive or insufficient for the current load torque. Human users can utilize the feedback information from the biological sensors in fingertips to determine the actual load torque and the time at which users voluntarily alter their grip force is very consistent, which is about 0.11s. We also observed the distortion of the human hand movement trajectories in part due to the excessive or insufficient grip force. We have provided a detailed procedure to identify such a model using time-domain system identification methods. Based on this model, we are able to describe the distortion in the human hand trajectories due to unexpected load torque changes. The proposed model, together with identification methodology, may be used to explain other human movement principles. Having formulated the grasp and twist task as an input determination problem, we are in position to address other simple manual tasks, describing them with a simple model that describes how humans build motor plans that account for the coupled dynamics between the body and task object.
Submitted by James Freudenberg
on
Abstract:
When mechanical linkages are replaced by electronic communication and control systems, certain undesirable phenomena can arise that may be difficult to anticipate. In simple cases the lack of dissipativity in the cyber link unmasks instabilities that were present but suppressed in the system with a mechanical (physical) link. In more complex cases, there may be no simple physical equivalent to the system containing the cyber link, and a damping coefficient or dissipative element may not be identifiable. The major goal of this project is to develop appropriate models and accompanying control theory to describe the phenomena that arise in cyber-physical systems that do not have simple correlates in purely physical systems. The “cyber” system of particular interest for us in this work is a system containing a human in the loop. We have been calling these “cyber-physiological” systems, to highlight the role that biomechanics plays in system behavior. But even further, the physiological part of such systems contains control loops implemented in very complex and not-to-be-deciphered neural circuits. We seek appropriate models that capture phenomena observable in cyber-physiological systems, especially those related to the realization of certain “links” in neural circuits, certain links in electronic circuits (and embedded control code), and yet other links in physical components. To reduce the problem to a tractable form, we make certain assumptions based on empirical evidence concerning the timing of feedback control and the mix of feedback and feedforward processes within the human central and peripheral nervous system. We apply system identification techniques and modeling principles guided by control to obtain models which are as simple as possible yet competent to capture empirically observed phenomena. We are currently using a prototypical manual task, the grasp and twist task, which involves a significant amount of coordination that produces interesting observable phenomena under perturbations. In terms of the grasp and twist task, our goals are to capture all the necessary aspects of motion planning, force control, biomechanics, and appropriate mixes of feedforward and feedback control in a simple model. The grasp and twist task is designed using our single-axis rotary haptic device involving a programmable load force. The load force may be programmed to behave like a weight, spring, or damper. We also outfitted the device handle with a force sensor to measure grip force. We have developed a simple model of the human user incorporating a motion source to capture user intent (the neural command) and a spring-damper coupler with mass to capture the mechanics of muscle, tendon, and soft tissues. We use independent system identification experiments to identify the mass, stiffness, and damping parameters of this model. These experiments are independent in that they are performed prior to the grasp and twist task. We have determined, using empirical results and a study of model properties, that it is necessary to use mass, stiffness, and damping parameters that depend on a variable grasp force to capture the observed behaviors using our model of limb mechanics. That is, it is not possible to obtain a motion source trajectory that produces the observed behaviors with parameters that do not depend on grip force. Naturally, the grip force does change during the grasp and twist task and is available through our force sensor. From the data in cyber grasp and twist tasks, we found the fine coordination between load torques and grip forces is preserved. Human users can generate appropriate grip force to achieve the desired movement despite different load torques. This suggests that human users build an internal model of the external environment as well as the limb dynamics. From the experiment data of the grasp and twist task with unexpected load changes, we found users will prepare for the current trial based on the experience of the previous trial. Hence if the load torque changes, users will erroneously prepare for the previous load torque and hence the grip force will be excessive or insufficient for the current load torque. Human users can utilize the feedback information from the biological sensors in fingertips to determine the actual load torque and the time at which users voluntarily alter their grip force is very consistent, which is about 0.11s. We also observed the distortion of the human hand movement trajectories in part due to the excessive or insufficient grip force. We have provided a detailed procedure to identify such a model using time-domain system identification methods. Based on this model, we are able to describe the distortion in the human hand trajectories due to unexpected load torque changes. The proposed model, together with identification methodology, may be used to explain other human movement principles. Having formulated the grasp and twist task as an input determination problem, we are in position to address other simple manual tasks, describing them with a simple model that describes how humans build motor plans that account for the coupled dynamics between the body and task object.