Subtask sequence for the robot interface. The robot hand is controlled to grasp two clothespins and a ball. Each object is arranged such that the hand must change both position and orientation to grasp the object. The object is then placed into the bin below the table. The order in which these tasks are completed is determined by each subject. The clothespins must be grasped as shown in the images to successfully complete the task. 

Context in source publication

Context 1

... is an element-wise matrix multiplication, u ( ) is the unit step function, σ is the muscle activation threshold, and g is the output gain. Both σ and g can be tuned for each subject, and M is a semi-random mixing matrix converting F ˆ ( t ) to the control outputs U ( t ) . U ( t ) is averaged over the last five outputs to provide consistent control. In this experiment, the 7-DoF control scheme is imple- mented as a two-state finite state machine (FSM), with each state offering simultaneous control of velocities over 4-DoFs (see Table I). M is designed to cover the entire output space ( c = 4 ) using minimal inputs ( k = 6 ) while decoupling control axes 1-3 from control axis 4 (see Fig. 1): (4) State switching is done by monitoring the simultaneous threshold breech between the last two activation inputs, F ˆ 5 and F ˆ 6 , contributed by an antagonistic muscle pair. 1) Pre-Processing: The HDsEMG signals are subtracted from the mean of all channels to dampen the influence of common noise, and then rectified and low-pass filtered (fourth-order zero-lag Butterworth, cut-off 3Hz). Finally, the signals are filtered by a 3x3 median filter to minimize the effects of electrode lift-off. The sEMG signals of an additional antagonistic muscle pair are rectified, low-pass filtered (fourth-order zero-lag Butterworth, cut-off 3Hz), and normalized with respect to the subject’s maximal voluntary contraction (MVC) for these two muscles, as found during the initial calibration. Both series of signals are then sub- sampled to 200 Hz and merged to create Y ( t ) . 2) Calibration: Each subject is first guided through the calibration stage described in [20] to generate a unique W . A total of 16 wrist and finger motions from the right arm are investigated - wrist flexion/extension, wrist prona- tion/supination, ulnar/radial deviation, hand open/close and flexion/extension of the index, middle, ring, and pinky fingers. 192 HDsEMG signals are collected from the subject’s forearm using HD electrode grids to form an initial W ˆ 0 with k 0 = 4 . Two additional columns are added with unit input on the 193 rd and 194 th rows, respectively, and zeros elsewhere. These two columns contain the sEMG from biceps brachii (BB) and triceps brachii (TB), resulting in a 194 × 6 matrix W ˆ . During this calibration phase, subjects are also asked to perform their MVC for BB and TB to initially set the state switching threshold at 50% of it. MVC values are not needed from the HDsEMG, as explained in [10]. 3) Robot Control: There is a slight difference in operation between VR and robot control induced by joint limits, singularities, and inertia. LWR 4 operates in Cartesian impedance control using inverse kinematics when the control state is in position, and joint impedance control using forward kinematics of the three wrist joints when the control state is in orientation mode. The switch is enforced to reduce the risk of joint velocity and position limits being exceeded while rotating through singularities. Global ρ , φ , and θ are limited to ± π radians to avoid physical limitations while rotating. 3 The iLIMB operates via Bluetooth with velocity commands sent to open/close all fingers at 200 Hz . Two sEMG systems were used for data collection. The first system included 192 monopolar channels from the subject’s forearm using three equidistant semi-disposable adhesive 8 × 8 grids with 10 mm inter-electrode distance. The EMG- USB2, OT Bioelettronica amplifier was set to gain of 1000 with internal bandpass filter at 3 − 900 Hz , broadcasting samples via TCP at 2048 Hz with 12-bit depth for further processing, as in [18]. The second system included two bipolar channels placed on the BB and TB muscles. These wireless sEMG electrodes (Delsys Trigno Wireless) were acquired with a gain of 500, digitized with 16-bit depth at a frequency of 1926 Hz and broadcast via TCP. Both interfaces receive commands at 200 Hz from a custom program using C++ and OpenGL API [21]. This program performs real- time processing and conversion of sEMG inputs into control variables of linear velocity, angular velocity and color/grasp. The full setups are shown in Fig. 2 and 3, respectively. Subjects, without prior knowledge on how to control the interface, attended three sessions across several days. The first session consisted of the calibration phase described above, followed by an introductory control phase. The control phase introduced subjects to the VR helicopter, with 20 minutes of exploration, in which the subject was encouraged to explore the space and become familiar with the control paradigm, followed by 26 tasks to be completed. The tasks are distributed as to cover the entire volume of the task-space and require activation of all available DoFs, as explained in Fig. 4. After completing each full task, the helicopter returns to the center of the screen with an initial orientation and color followed by a ten second break. There was no time limit imposed in order to encourage users to explore and discover a comfortable control. The random arrangement of targets was consistent for each subject in the experiment. The second session occurred at least 24 hours after the first. Subjects were given one hour to accomplish as many tasks as possible while using the same control scheme and W ˆ calculated during the first session. This session provided data regarding learning retention and continued learning trends. The final session occurred between one and eight days after the second. Subjects were introduced to the robot myoelectric interface, while using the same control scheme and W ˆ calculated in session one. Subjects are asked to complete three precision tasks, with no strict order, by sequentially grasping a tennis-sized ball and two customized clothespins to place in a bin. The task sequence is timed and shown in Fig. 5. This session provided evidence of precision control capabilities and learning transfer despite slightly different system dynamics of the robot compared to the VR. During the first two sessions, collected datasets contained values describing task difficulty, completion times, and path lengths used to accomplish each task. This data was analyzed in data blocks containing 25% of each session’s data from all subjects. The total completion time is recorded for the third session to indicate precision performance capabilities and any factors influencing these capabilities. 1) Learning Trends: Metrics used for assessing performance in the first two sessions are provided in Table II, using first degree polynomials to fit the results with respect to block number. These linear trends are assumed according to [10], as the initial exponential learning component has been accounted for in the first 20 minutes of exploration. CT is the time needed to fulfill the task [22]. T P , ex- pressed in bits/second according to Fitts’ law [23], measures both speed and accuracy by considering the difficulty of the task [9]. P E is the ratio between the shortest path possible to complete the entire task and the actual path taken to reach the target [24]. b is the overall block number in session 1 and 2, κ is initial performance, and β shows the learning rate, such that β > 0 indicates better performance and a significant learning component, for each metric. The index of difficulty, ID , of a given task is given by the Shannon Formulation ...