Fig 1.
Exoskeleton systems used in the experiments, developed at BRI ATR.
The top photo shows the elbow exoskeleton [34] and the bottom photo shows the whole-arm exoskeleton [35].
Fig 2.
Schematic presentation of proposed adaptive exoskeleton control system.
The human is included into the robot control loop through EMG-based biofeedback and DMP learning system. DMP system was made to minimise the EMG-based biofeedback by generating appropriate periodic feed-forward exoskeleton control trajectories. The adaptive oscillators are used to extract the phase and frequency information from human biofeedback and modulate the phase and frequency of the learnt feed-forward trajectories.
Fig 3.
Results of experiment on elbow exoskeleton using a trained subject.
The first graph shows commanded PAM pressure as it was learned by the DMP system. The second graph shows the joint torque calculated based on the force measurement from the load cell. The third graph shows the human muscle activity based feedback obtained from the measured EMG. The fourth graph shows the angle of the robot elbow joint as measured by the encoder in the exoskeleton joint. The fifth graph shows the frequency of motion as extracted by the adaptive oscillator from the muscle activity feedback U. The given referent states are highlighted by black lines.
Fig 4.
Sequence of photographs taken during the experiment that illustrate the state of the exoskeleton at the end of the 3 stages.
Images in the first sub-row show the exoskeleton when it moved the object to the lower reference position, while the images in second sub-row show it in the upper position.
Fig 5.
Control pressure trajectory updates as a function of phase.
The graphs show trajectories as they were gradually learnt by the DMP system and fed to PAM pressure controller. Each colour variation represents one periodic cycle. The time sequence is illustrated by the colour spectrum, where the earlier trajectories are plotted with red and the later with violet colour. The left graph shows the first stage, the middle graph the second stage and the right graph shows the third stage of the experiment.
Fig 6.
The three graphs in the first row show the end-effector motion at the end of there different stages of experiment (4 phases plotted for each).
The red colour represents the earlier while violet represents the later trajectories. The lower two graphs show the control pressure trajectories updated as a function of phase for each joint.
Fig 7.
Results of experiment on whole-arm exoskeleton using a trained subject.
The upper four graphs show the variables related to the shoulder joint. The lower five graphs show the variables related to the elbow joint. The last graph shows the estimated frequency from the elbow joint muscle activity feedback. Variable p corresponds to the commanded pressure learned by the DMP system, variable τ to the joint torques calculated from measured forces in the load cells, variables US and UE to muscle activity feedback signals from shoulder and elbow flexion/extension joints, variable Θ to measured joint angles by the encoders and Ω to the frequency estimated by the adaptive oscillator from UE.
Fig 8.
Sequence of photographs taken during the experiment that illustrate the state of the exoskeleton at the end of the 3 stages.
Images in the first sub-row show the end-effector when it moved the object to the lower reference position, while the images in second sub-row show it in the upper position.
Fig 9.
Results of the experiment on multiple subjects.
Graphs are showing normalised tracking error of amplitude and frequency with respect to the given reference conditions in passive and active mode. Standard error of mean is marked with black.
Table 1.
Conceptual comparison between different types of control methods.