As a contribution toward the goal of adaptable, intelligent artificial limbs, this work introduces a continuous actor-critic reinforcement learning method for optimizing the control of multi-function myoelectric devices. Using a simulated upper-arm robotic prosthesis, we demonstrate how it is possible to derive successful limb controllers from myoelectric data using only a sparse human-delivered training signal, without requiring detailed knowledge about the task domain. This reinforcement-based machine learning framework is well suited for use by both patients and clinical staff, and may be easily adapted to different application domains and the needs of individual amputees. To our knowledge, this is the first my-oelectric control approach that facilitates the online learning of new amputee-specific motions based only on a one-dimensional (scalar) feedback signal provided by the user of the prosthesis.

Objective: Persons with normal arm function can perform complex wrist and hand movements over a wide range of limb positions. However, for those with transradial amputation who use myoelectric prostheses, control across multiple limb positions can be challenging, frustrating, and can increase the likelihood of device abandonment. In response, the goal of this research was to investigate convolutional neural network (RCNN)-based position-aware myoelectric prosthesis control strategies. Methods: Surface electromyographic (EMG) and inertial measurement unit (IMU) signals, obtained from 16 non-disabled participants wearing two Myo armbands, served as inputs to RCNN classification and regression models. Such models predicted movements (wrist flexion/extension and forearm pronation/supination), based on a multi-limb-position training routine. RCNN classifiers and RCNN regressors were compared to linear discriminant analysis (LDA) classifiers and support vector regression (SVR) regressors, respectively. Outcomes were examined to determine whether RCNN-based control strategies could yield accurate movement predictions, while using the fewest number of available Myo armband data streams. Results: An RCNN classifier (trained with forearm EMG data, and forearm and upper arm IMU data) predicted movements with 99.00% accuracy (versus the LDA's 97.67%). An RCNN regressor (trained with forearm EMG and IMU data) predicted movements with R ² values of 84.93% for wrist flexion/extension and 84.97% for forearm pronation/supination (versus the SVR's 77.26% and 60.73%, respectively). The control strategies that employed these models required fewer than all available data streams. Conclusion: RCNN-based control strategies offer novel means of mitigating limb position challenges. Significance: This research furthers the development of improved position-aware myoelectric prosthesis control.

The relationship between a reinforcement learning (RL) agent and an asynchronous environment is often ignored. Frequently used models of the interaction between an agent and its environment, such as Markov Decision Processes (MDP) or Semi-Markov Decision Processes (SMDP), do not capture the fact that, in an asynchronous environment, the state of the environment may change during computation performed by the agent. In an asynchronous environment, minimizing reaction time—the time it takes for an agent to react to an observation—also minimizes the time in which the state of the environment may change following observation. In many environments, the reaction time of an agent directly impacts task performance by permitting the environment to transition into either an undesirable terminal state or a state where performing the chosen action is inappropriate. We propose a class of reactive reinforcement learning algorithms that address this problem of asynchronous environments by immediately acting after observing new state information. We compare a reactive SARSA learning algorithm with the conventional SARSA learning algorithm on two asynchronous robotic tasks (emergency stopping and impact prevention), and show that the reactive RL algorithm reduces the reaction time of the agent by approximately the duration of the algorithm's learning update. This new class of reactive algorithms may facilitate safer control and faster decision making without any change to standard learning guarantees.

Successful hand-object interactions require precise hand-eye coordination with continual movement adjustments. Quantitative measurement of this visuomotor behaviour could provide valuable insight into upper limb impairments. The Gaze and Movement Assessment (GaMA) was developed to provide protocols for simultaneous motion capture and eye tracking during the administration of two functional tasks, along with data analysis methods to generate standard measures of visuomotor behaviour. The objective of this study was to investigate the reproducibility of the GaMA protocol across two independent groups of non-disabled participants, with different raters using different motion capture and eye tracking technology. Twenty non-disabled adults performed the Pasta Box Task and the Cup Transfer Task. Upper body and eye movements were recorded using motion capture and eye tracking, respectively. Measures of hand movement, angular joint kinematics, and eye gaze were compared to those from a different sample of twenty non-disabled adults who had previously performed the same protocol with different technology, rater and site. Participants took longer to perform the tasks versus those from the earlier study, although the relative time of each movement phase was similar. Measures that were dissimilar between the groups included hand distances travelled, hand trajectories, number of movement units, eye latencies, and peak angular velocities. Similarities included all hand velocity and grip aperture measures, eye fixations, and most peak joint angle and range of motion measures. The reproducibility of GaMA was confirmed by this study, despite a few differences introduced by learning effects, task demonstration variation, and limitations of the kinematic model. GaMA accurately quantifies the typical behaviours of a non-disabled population, producing precise quantitative measures of hand function, trunk and angular joint kinematics, and associated visuomotor behaviour. This work advances the consideration for use of GaMA in populations with upper limb sensorimotor impairment.