Latent Variable Grasp Prediction for Exoskeletal Glove Control

Abstract

This paper presents a grasp prediction algorithm designed to govern the motion of an exoskeletal glove in rehabilitative and assistive applications. Recent research into the dynamics of hand motion has shown that the complex motion of the finger joints can be represented as a smaller set of coordinated motions or latent variables. This fact forms the basis of the proposed algorithm capable of successful prediction even with noisy data. From relatively small motion (minute user hand movements) as the input, the developed algorithm can predict intended grasp configurations. The 16 finger joint angles, with random noise, are mapped onto a set of six latent variables for which the estimated noise and future configuration are simultaneously determined using a linear regression. The algorithm was tested in simulation on published motion data from 30 healthy subjects performing a set of common grasps on multiple objects. The algorithm was able to determine the target state with an accuracy of approximately 90% for each subject, despite the nonlinear motion and non-uniform trajectory variations. We propose that the predicted grasp is an adequate target for an exoskeletal glove to provide initial gross movement for the user, then iteratively converge to the desired grasp with only limited additional user input.

Grasp Prediction Toward Naturalistic Exoskeleton Glove Control

“…The effects on the accuracy and speed of varying the threshold parameter μ can be seen in [24] where a larger threshold required an increased percentage of the grasp to be completed before (9) converges. Similar effects are seen for varying α, σ, and γ.…”

“…For many grasps, it can be seen that the regression-based prediction never predicts the correct grasp. As discussed in [24], while it is acceptable to converge to a similar grasp, continued iterations of the prediction as more of the grasp is completed is likely to converge to the absolute intended grasp. The trajectory-based prediction is superior in this regard, as only a single run of the prediction, lasting less than one quarter of the grasp motion, results in this absolute determination.…”

“…These prediction schemes can be used in a haptic controller to assist the user of an exoskeleton glove when they are unable to successfully maneuver their hands of their own volition. Section II of this paper will present our existing regression-based prediction algorithm [24] and our new probabilistic comparison algorithm while highlighting the improvements in performance shown by the newer approach. Section III will discuss the experiments conducted using the RML Glove and the results of both the simulations and the experiment.…”

A Series Elastic Actuator Design and Control in a Linkage Based Hand Exoskeleton

“…Additionally, more ranges and types of motion will be tested such as small "holding" motions and following a trajectory. The actuator and controller will be integrated into a full prototype, along with a holistic grasp controller that uses the amplified motion to predict the grasping action the user is attempting [22]. The complete design will be made as light as possible to ensure portability and usability.…”

“…The goal of this work is to provide an actuator design and control structure that facilitate improved gross hand motion by an impaired individual. Prediction algorithms, as shown in [22] could be used to analyze the motion and assist the user with more specific grasping actions. In this way, we hope to accomplish both general and unique finger motion to assist the wearer fully through the grasping motion.…”