Human-exoskeleton interaction portrait

Abstract

Human-robot physical interaction contains crucial information for optimizing user experience, enhancing robot performance, and objectively assessing user adaptation. This study introduces a new method to evaluate human-robot interaction and co-adaptation in lower limb exoskeletons by analyzing muscle activity and interaction torque as a two-dimensional random variable. We introduce the interaction portrait (IP), which visualizes this variable's distribution in polar coordinates. We applied IP to compare a recently developed hybrid torque controller (HTC) based on kinematic state feedback and a novel adaptive model-based torque controller (AMTC) with online learning, proposed herein, against a time-based controller (TBC) during treadmill walking at varying speeds. Compared to TBC, both HTC and AMTC significantly lower users' normalized oxygen uptake, suggesting enhanced user-exoskeleton coordination. IP analysis reveals that this improvement stems from two distinct co-adaptation strategies, unidentifiable by traditional muscle activity or interaction torque analyses alone. HTC encourages users to yield control to the exoskeleton, decreasing overall muscular effort but increasing interaction torque, as the exoskeleton compensates for user dynamics. Conversely, AMTC promotes user engagement through increased muscular effort and reduces interaction torques, aligning it more closely with rehabilitation and gait training applications. IP phase evolution provides insight into each user's interaction strategy formation, showcasing IP analysis's potential in comparing and designing novel controllers to optimize human-robot interaction in wearable robots.

Keywords: Control; Exoskeleton; Physical Interaction.

Fig. 1

Regions of interaction portrait (IP). Each quadrant of the circle corresponds to different human–exoskeleton interaction modes determined by the variation of the normalized total muscle activation ($_{c_1}^{c_2}\mathrm{\Delta }\mu$) with respect to the normalized total interaction torque ($_{c_1}^{c_2}\mathrm{\Delta }\tau$) between controllers $c_1$ and $c_2$, respectively. The first quadrant (red) indicates increased disagreement between the user and exoskeleton, resulting in an increase in both muscle effort and the total interaction torque. The second quadrant (green) determines the co-adaptation of the user toward participating in the motion as much as possible and leading the motion. The third quadrant (blue) denotes the decrease in total interaction torque and the total muscle effort, associated with the decrease in human–exoskeleton disagreement. Finally, the fourth quadrant (orange) denotes the condition at which the user yields control of the motion to the exoskeleton and minimally activates their muscles. In this case, muscle activation decreases while the total interaction increases since the exoskeleton has to carry the user’s body (passive dynamics) in addition to the exoskeleton dynamics

Fig. 2

Block diagram of TBC, HTC, and AMTC controllers. A Block diagram of the Time-Based Controller (TBC). A time-based gait phase along with the estimated gait speed are fed into a lookup table to determine the applied torque to the exoskeleton joints according to joint torque data recorded from the exoskeleton during high-gain joint control with the user passively following the exoskeleton (with the minimum voluntary contribution to the gait). B Hybrid Torque Controller (HTC) consisted of a data-driven estimator of the required joint torque along with a lookup table-based torque controller similar to the TBC. In this case, however, the gait phase is determined according to the exoskeleton states rather than time. The torque from the two different pipelines is finally combined with the weight of $w=0.75$ and $1-w=0.25$ to form the applied torque to the exoskeleton. C Block diagram of the Model-Based Torque Controller (AMTC). The gait phase is estimated according to the exoskeleton joint angles and then fed into a trajectory adaptation block which learns the joint trajectory of the participant in real time and uses that trajectory as the reference for the exoskeleton to be fed into the forward dynamics of the exoskeleton to determine the feedforward joint torques

Fig. 3

A Dorsal, lateral, and frontal view of a participant with the Indego exoskeleton with active hip and knee joints. The participant is standing on the Bertec treadmill with two speed-controlled belts equipped with individual loadcells underneath each of them for GRF monitoring. Muscle activation is measured from both right and left leg muscles using EMG sensors. Gait up IMUs are clipped to the outer side of the shoes, right below ankle joints to measure the spatiotemporal parameters of gait. Oxygen uptake of the participant is measured and recorded at each breath through a mask connected to the gas analyzer carried at the back of the participant. B Treadmill speed changes while experimenting with the HTC, AMTC, and the TBC controllers. The order of the controllers was specific to participant #9 and varied for other participants. Participants walked with each controller for 300 s divided by three 100-second walking periods during each the treadmill speed was set to 0.4, 0.6, and 0.8 m/s, respectively

Fig. 4

Examples of a typical participant’s (#9) experimental data; for ease of visualization and interpretation, only the interaction torque at the right hip and activation of one of the muscles are illustrated here along with the relative oxygen uptake. A The mean absolute interaction torque at the right hip at each stride with each controller and speed for Participant #9. B Normal muscle activation for the Gastrocnemius Medialis (GM) at the right leg. The dashed line represents average activation during no-exoskeleton walking. GM was chosen since it showed the strongest sensitivity to changes in the controller. C Relative oxygen uptake with respect to no exoskeleton walking for each breath for each controller and speed. The oxygen uptake has increased with the increase in treadmill speed

Fig. 5

Muscle activation, interaction torque, and exoskeleton applied torque profiles with respect to the gait phase for participant #9. A The average normalized muscle activation pattern for the TBC, HTC, and AMTC blocks for the right Gastrocnemius Medialis during walking at 0.4, 0.6, and 0.8 m/s. The shaded area represents the standard deviation of the muscle activation about their mean value. Similarly, the average human–exoskeleton interaction torque and exoskeleton joint torques at the right hip are plotted in B and C, respectively

Fig. 6

The average performance metrics for each treadmill speed and controller across participants. A The sum of the relative oxygen uptake across all the strides for each speed in each controller block graphed for each participant. The bars show the average of the sum of the oxygen uptake across all participants. Similarly, the average total absolute value of the human–exoskeleton interaction and total normalized muscle effort are graphed in B and C, respectively. Asterisks indicate statistical difference between the median of the compared populations

Fig. 7

Comparing the average interaction portrait for each pair of controllers. The average interaction portrait (IP) depicted according to the average total muscle effort and the average total human–exoskeleton interaction for each participant computed at each of the 0.4, 0.6, and 0.8 m/s speeds for the TBC$\to$HTC, TBC$\to$AMTC, and HTC$\to$AMTC illustrated in A, B, and C, respectively. The yellow areas denote the area between 25 and 75 percentiles

Fig. 8

Interaction portrait distribution along with their polar histogram for HTC and AMTC blocks with respect to the average total muscle effort and total interaction torque across all strides during the TBC block graphed for each participant plotted for walking at 0.8 m/s. The radial coordinate of data points is normalized with respect to the maximum radius computed across all participants’ strides. Participants are arranged increasingly according to their body mass. The polar histograms show the concentration intensity of the depicted points. Each bin of the histogram covers $\pi /6$ rad

Fig. 9

Evolution of IP phase at each stride at different walking speeds for two sample participants. The top and bottom rows depict IP phase evolution during walking at 0.4, 0.6, and 0.8 m/s for each of the TBC$\to$HTC and TBC$\to$AMTC cases for participants #5 and #4, respectively

Fig. 10

A Evolution of the modification term for Participant #9’s right hip trajectory (${\mathrm{\Delta }}_{\text{hip}}$). ${\mathrm{\Delta }}_{\text{hip}}$ is zero before and after the AMTC block. During the AMTC block, the modification term of the trajectory adapts to make the reference trajectory closer to the user joint angle. At each walking speed, the modification pattern in the steady state converges to a different amplitude and shape, indicating that the user joint angle, and as a result, the exoskeleton reference trajectory has evolved to a different pattern at each speed. B Modification trajectory coefficients evolve during the AMTC block. Each coefficient converged to a steady state value at the end of each treadmill speed condition. After the change in treadmill speed, coefficients converge to new optimum levels. These levels, as mentioned for the modification term of the hip trajectory, differ for each speed. C Pearson correlation between the GRF of each stride with the average GRF profile recorded during natural walking without the exoskeleton. D Stride length was computed using the Physiolog 6 s IMU sensors

Fig. 11

The average maximum heel clearance (A), minimum toe clearance (B), stance percentage (C), and stride length (D), computed for each of the 0.4, 0.6, and 0.8 m/s speeds during natural walking with no exoskeleton, TBC, HTC, and the AMTC blocks for each participant. Bars show the average of each metric across participants. Asterisks indicate statistical difference between the median of the compared populations

Fig. 12

A The average Pearson correlation computed between the average vertical GRF in natural walking without the exoskeleton and the vertical GRF during walking with each of the proposed controllers at different speeds. The depicted bars show the average of each metric across participants. B and C shows the similar graph for the mediolateral and antero-posterior GRF plotted only for walking at 0.8 m/s. Astrisks indicate statistical difference between the median of the compared populations