Abstract
Purpose
User preference is important for improving the acceptance of assistive robotic devices. However, it is not clear what factors influence user preferences or how it relates with physiological variables, such as muscle activity.
Methods
Three participants with transfemoral amputation walked on a treadmill at a self-selected speed using a knee-ankle prosthesis. In four trials conducted on the same day, participants identified their preferred push-off parameters by self-tuning the magnitude and timing of the assistance using a 2D grid interface through a self-exploration method. They were blinded to the control parameters and relied on their perception of the assistance to guide the tuning. During these trials, muscle activity was recorded from four muscles in the sound limb. Muscle intensity and muscle synergies were calculated to explore the relation between muscle activity and user preference.
Results
While the participants generally preferred low to medium magnitudes of assistance, their preferred settings were not consistent across trials. In most trials these preferred settings were associated with lower muscle intensity in most muscles and better motor coordination.
Conclusion
These findings suggest that physiological metrics, such as muscle intensity and muscle synergies, may influence the selection of preferred prosthesis settings. Even though user preference is likely influenced by multiple factors, such as gait symmetry, understanding the role of muscle activity and its relationship with preference could support the development of more personalized assistive robotic devices. However, further research is needed to confirm these observations and better explain how muscle activity contributes to participants’ preferred settings.
Supplementary Information
The online version contains supplementary material available at 10.1186/s12984-026-01931-w.
Keywords: Prosthetics, User preference, Human-in-the-loop optimization
Introduction
Wearable robots are devices designed around the shape and function of human limbs, providing additional forces that can overcome physical limitations [1] and enhance human mobility. For example, prosthetic devices aim to replace the limb after amputation and provide mobility benefits for amputees. However, passive lower-limb prostheses lack the ability to deliver the mechanical energy typically provided by biological muscles [2]. To address this, researchers are advancing to powered prostheses capable of replicating able-bodied kinematics and kinetics by using electric motors that add net-positive mechanical energy, analogous to the muscles within a leg. In addition, the field of assistive robotics has made significant progress in customizing control strategies to optimize performance for specific users and tasks. A major challenge, however is adjusting the device’s behavior to meet the unique needs of each individual. Therefore, recent methodologies, such as human-in-the-loop optimization (HILO), employ objective-driven functions to fine-tune control parameters [3]. While these approaches have shown promising results, most optimization efforts focus on a single physiological (e.g., metabolic cost) [4–8], or biomechanical (e.g., gait parameters) [9–11] objective functions, often overlooking the user’s subjective experience. In reality, there are many factors that influence the user’s experience which sometimes cannot be captured by the available approaches, and at the end user experience is a key concept for users to actually use any wearable assistive robot.
To address this, researchers have started incorporating user preference as an objective function in the optimization of wearable robotic systems. Preference is a complex and powerful variable that reflects users’ needs, desires, and perceptions [12]. It reflects a combination of factors, as users can perceive context-related objectives (e.g., speed, effort) and subjective experiences (e.g., comfort) [12, 13], bridging the gap between performance and perception. However, since preference is inherently a multi-objective function, it is a challenge to define it as a single cost function. Since, the physiological and biomechanical factors that shape user preference remain unclear. Then, understanding how preference relates to measurable outcomes is essential for developing effective and personalized control strategies.
The selection of suitable objective functions is essential for achieving optimal performance in wearable robotic systems [12]. To effectively use user preference in optimizing lower-limb prostheses, it must be properly assessed. In addition, it is important to understand how preference evolves as users adapt to robotic assistance. Recent studies have explored preference-based optimization in lower-limb [14–16] and upper-limb [17] exoskeletons, as well as prosthetic devices [18]. There have been several ways to measure preference, Ingraham et al. [12] proposed a comprehensive framework for modeling and assessing user preference, outlining different methods used to evaluate it. Their work also highlights various strategies for exploring the preference landscape, including user-driven exploration [14, 19] and algorithm-based querying approaches [13, 18].
Some researchers have studied how user preference is related to physiological or biomechanical variables. For example, Clites et al. [15] reported that preferred ankle stiffness optimized ankle kinematic symmetry but did not minimize metabolic cost. Other biomechanical descriptors showed significant trends but did not show an optimal result at the preferred stiffness. Based on this idea, a data-driven model was proposed to predict user preference using biomechanical data [20]. However, the relationship between user preference and muscle activity, a physiological measure linked to effort and metabolic cost [21], remains unexplored.
Therefore, the main goal of this study is to personalize the ankle prosthesis assistance by allowing users to select two control parameters using a self-exploration grid that shapes the assistance profile. In addition, the study aims to provide insights into physiological factors, extracted from muscle activity signals, that may explain what guides user preference. We hypothesize that muscle activity can offer metrics related to motor efficiency and perceived exertion. This hypothesis is supported by trends observed in previous studies, which report increased motion variability with changes in push-off timing. Our work demonstrates that analyzing muscle activity can help identify physiological metrics linked to user preference. Understanding these links is an important step toward integrating user preference into the design and evaluation of assistive technologies.
Methods
The prosthetic device used in our experiments is the same device we used to compete in the CYBATHLON 2024. We participated in the leg prosthesis discipline, which involved performing everyday tasks such as ascending a staircase while balancing a cup, walking up a ramp, or stepping on top of wobbly stones. Beyond being a competition, CYBATHLON demonstrates the importance of studying biomechanics to support the development of assistive robotic devices. Throughout our preparation and participation, it became clear that the controller assistance must be personalized to each user’s comfort and preference. Understanding the underlying factors that influenced our pilot’s choices was essential to effectively adjust the controllers. These user preferences often reflect a combination of physiological and biomechanical considerations that are not captured by performance metrics alone. This experience was a major motivation to continue working on the personalization of assistance and to further improve the knee-ankle prosthetic device.
Preference experimental protocol
Three participants with a transfemoral amputation (age =
years, body mass =
kg, height =
cm) walked on a treadmill at a self-selected speed (Additional file 1: Table S1) using the prosthesis presented in Sect. 2.2. They were recruited from an existing participant network maintained by our research team. All participants had good functional mobility (K3 activity level or higher), no complex lower-extremity injuries, and were consistent users of conventional passive prostheses. Two participants used passive knee-ankle prostheses for daily activities and one used a microprocessor-controlled prosthesis. Only one participant had previous experience with an active device, specifically from earlier testing sessions of our knee-ankle prosthesis (Additional file 1: Table S1).
Using a touch-screen interface, they adjusted the timing and magnitude of the ankle prosthesis plantar flexion torque in real time to find their preferred push-off settings. This approach is known as the self-exploration paradigm, where the user adjusts the behavior of the device, and decides whether or not they have reached their preference [12]. Participants performed four preference-identification trials. During each trial, participants began with a random grid from the four possible cases (described in more detail in Sect. 2.4) and a randomized initial condition. The participants were instructed to explore the grid to find their preferred settings and to experience each set of parameters for a maximum of 20 s. As soon as they found their preferred settings, they informed the experimenter, who stopped the trial. We verbally reminded participants to explore the space until they were confident in their selection, but there was no limitation on the number of parameters that could be tested (Additional file 1: Fig. S1).
Before the testing session, each participant completed a separate familiarization session that lasted approximately three hours. During this time, the prosthesis was fitted to the participant and the knee controller parameters were manually adjusted to ensure comfort (see Sect. 2.3). The ankle joint remained passive during this time. The participant walked in multiple short walking bouts until they felt comfortable with the knee settings. The participant then completed a practice trial to become familiar with the ankle behavior and the touchscreen interface on the treadmill. During this trial, they were encouraged to explore the full range of control settings and informed that some combinations might feel uncomfortable.
We placed an IMU on the thigh of each participant to estimate the gait phase, as explained in Sect. 2.3, and compute the angle and velocity of the thigh. The IMU’s used in this study are those described by Munih et al. [22], developed for real-time feedback in wearable robotics.
To monitor physiological responses during the trials, we placed surface electromyography (EMG) sensors on four muscles of the participant’s sound limb (Trigno surface electromyography system, Delsys Inc., Natick, MA, USA). The selected muscles were the Tibialis Anterior (TA), Gastrocnemius Lateralis (GL), Gastrocnemius Medialis (GM), and Rectus Femoris (RF). These muscles were chosen to capture lower-limb activity relevant to gait dynamics [23, 24]. A reduced number of sensors was chosen to facilitate future translation of this approach to more realistic, real-world conditions.
Prosthesis hardware
Participants wore a transfemoral prosthesis, with active knee and active ankle modules designed at the Vrije Universiteit Brussel (Fig. 1A). The knee actuator is a 5-bar mechanism that provides torque through the actuation of a ballscrew (Misumi, Tokyo, Japan) by a brushless motor (EC 4-Pole 200W, Maxon, Sachseln, Switzerland). The non-linear effective gear ratio of the knee output was empirically determined through the use of a load sensing test bench over the range of motion of the knee (0 deg extension to 110 deg flexion). The knee measures the knee angle and velocity through the use of a magnetic absolute encoder and an incremental encoder placed on the motor. The prosthesis ankle (Fig. 1B) consists of the footplate, two fiberglass beam springs, ballscrew actuator (Misumi, Toyko, Japan), motor (EC 4-Pole 200W, Maxon, Sachseln, Switzerland), and shank. The footplate is connected to the shank through a single axis joint. The footplate is also connected to the two beam springs. The shank and beamsprings are connected through the motor/ballscrew actuator, with a pivot joint on each end. This beamspring has different contacts with the shank, the axle, and the footplate depending on the ankle angle and desired torque. A lookup table between the ankle angle (15 degrees dorsiflexion to 40 degrees plantar flexion), motor position, and output torque was created through the use of an actuated load sensing test bench. The ankle measures the ankle joint angle and velocity through the use of a magnetic absolute encoder and an incremental encoder placed on the motor. The knee housed all the computing units (sbRIO-9651, National Instruments, Austin TX) for both the knee and ankle motors, as well as provided communication with external devices and mid-level control of the prosthesis. The knee also contained the batteries (8x ICR18650, Samsung, Seoul, South Korea) for all the prosthesis power. Motor drivers (Twitter Gold, Elmo, Petach-Tikva, Israel) are embedded in the knee and ankle using custom interface boards.
Fig. 1.

Knee and ankle prosthesis used in this study
The knee and ankle used a standard pyramid adapter and was sized to participants using their socket. The complete prosthesis weighs 5.3 kg, in the standard length configuration, with foot cover and shoe. The device is synced with Inertial Measurement Units (IMUs), which are used to guide the finite-state machine (FSM) controllers.
Prosthesis control
The high-level control of the device, commanding its behavior, is obtained through a manual user input interface, with over 8 different selectable tasks. During each trial, the experimenter directly selected the walking task, reducing the cognitive load on the subject. Then the mid-level control of the walking task is defined with an FSM for knee actuation and a continuous phase variable approach for the ankle. An IMU needs to be placed on the participant’s thigh to guide the different controllers using the angular position and velocity of the thigh [25].
Knee control
The walking FSM was divided into four states: Idle, One Step, Stance, and Swing. The last two are the main states during steady walking, while Idle represents the standing position and One Step is used as a safety measure to avoid misfiring another state due to small movements when standing.
At the beginning of each trial, when the joint is fully extended, the position of the motor encoder is calculated and used as a reference for Idle, One Step and Stance, three gait states that use a position controller. In these states, data from the knee motor position is used to control the prosthetic knee. To provide continuous support and safety to the user, the target position is set to an overextended value (the reference) followed by a PID controller to ensure that the current at the knee joint was always greater than zero, keeping the knee fully extended.
During Swing, a feedforward square-wave torque is commanded at the knee. Since this state contributes to joint flexion, it is important for foot clearance and contributes to the forward progression of the leg, hence reducing the effort of the upper-leg and hip muscles.
During walking, the transitions between Stance and Swing are determined by a continuous phase variable (explained in Sect. 2.3.2) that describes the progression of the user through the gait cycle. The swing phase begins after toe-off, typically between 50 and 60% gait cycle, depending on the timing parameter selected by the user. The knee returns to full extension once the angular velocity of the thigh reaches zero (moment of maximum hip flexion). The knee is fully extended before the end of the gait cycle to ensure user safety and stability.
As mentioned earlier, the knee controller parameters were first manually adjusted until the participant felt comfortable. These parameters included the torque applied during the Swing and the thigh velocity used to return the knee to full extension before Stance.
The stance phase is important during walking, as it influences hip movement and helps reduce the braking forces on the intact limb [7, 26]. In this study, we controlled only the ankle joint during stance, while the knee was kept overextended to ensure stability. The primary objective was to demonstrate that, using the available hardware, participants could select their preferred assistance settings during push-off. Then, these user-selected settings were analyzed for their relationship with muscle activity. Keeping the knee overextended was not expected to impact these outcomes.
Ankle control
The ankle controller provides assistance as a function of the gait phase, and was controlled using a position-based strategy, where a desired motor position profile was defined by two shaping parameters: timing (percentage of the gait cycle) and magnitude (motor angle in revolutions). The commanded position profile was enforced on the device through a closed-loop controller. Note that the position profiles are referred to assistance profiles throughout this study. The resulting ankle torque was determined by the relationship of the commanded motor position, ankle kinematics, and the characteristics of the non-linear transmission and beam springs.
Torque was estimated using a lookup table constructed from bench tests, which related ankle angle and motor position to the output torque. At each instant, the ankle angle was measured using an encoder, and the corresponding torque was inferred from the lookup table using the commanded motor position. Although this approach allowed for straightforward implementation of parameterized assistance, it introduced variability in the resulting torque across participants and steps, since the final torque depended on the user’s gait and the prosthesis’ mechanical response.
Therefore, this study refers to the magnitude of assistance in terms of the commanded motor position, rather than torque, which was not directly controlled. In addition, we evaluated the overall effect of the commanded assistance profile on ankle kinematics, specifically focusing on the relationship between assistance magnitude and ankle plantar flexion following push-off (see Additional file 1: Fig. S2).
The assistance profiles were designed using two cubic splines between the desired set of parameters, assistance timing (
) and magnitude (
), and the onset and offset values selected at (30,0) and (70,0), respectively (as shown in Fig. 2 under assistance profile). To ensure continuity between the two splines, the first derivative is set to zero at the intersection between the two curves. Zero magnitude of assistance does not mean zero torque, and allows the ankle to behave as a passive device.
Fig. 2.
Participant self-tuned the ankle prosthesis controller using a blank two-dimensional grid displayed on a tablet. Selecting a space in the grid automatically changed the assistance profile that the participant experienced by changing the two parameters: magnitude and timing. Here we illustrate the mapping between the parameter coordinates and corresponding assistance profile. Solid black circles indicate the onset and offset values of the assistance profile
As mentioned above, the gait phase variable determines the changes of the FSM and the assistance profile of the ankle. The phase variable (
) is calculated using the approach proposed by Gregg et al. [27] using the thigh angle (
) and its integral (
). It is computed as follows:
![]() |
1 |
where
,
and z correspond to two shifting and one scaling factors, respectively. The first two center the range of values of the variables around zero while the latter (z) restores cohesion between the range of units of the thigh angle and its integral over time. These parameters are derived from the previous stride and they are computed as follows:
![]() |
2 |
Where
,
,
, and
are the maximum and minimum angle of the thigh, and the integral values of the previous stride. Since the output of Eq. 1 ranges between
and
, the phase is shifted and scaled between 0 and 100 to express the gait phase as a percentage of movement completion. The integral is reset at the start of each stride to prevent drift over time. This phase variable is then used as a reference for the generation of the ankle motor trajectory.
Interface
To allow participants to tune the ankle-prosthesis assistance settings in real-time, we used a touch screen tablet mounted on the treadmill, as shown in Fig. 2. The touch screen displayed a two-dimensional grid of selectable buttons. When a participant selects an element from the grid, its color changes to green to indicate the current setting; previous settings remained on the screen with a light gray color.
Selecting an element from the grid changed the assistance settings that the participant experienced, starting at the next heel contact with the prosthesis. Even though, the participants did not know what each grid element meant, the Cartesian coordinates of the grid were mapped to the parameter coordinates (timing and peak torque magnitude). To avoid participants returning to the same location (x, y) in each trial, we used four maps in which the x and y axes changed orientation, smallest to largest or largest to smallest with respect to the origin [14]. The map was randomized between trials and had no influence on the exploration strategy of the users.
For the timing parameter, the range of settings available to the participants was 46% to 56% gait cycle. This range was selected following previous research [14, 28]. Even though we informed the participants that they might experience uncomfortable assistance values, we did not aim to decrease their ability to walk normally. For the magnitude parameter, participants could explore 10 different levels of assistance, corresponding to motor angles ranging from 3 to 45 revolutions. The commanded motor position produced estimated torques ranging from 25 to 160 Nm, but, as discussed in Sect. 2.3.2, torque was depended on ankle kinematics. Since it was not a consistent or directly controlled variable, it was less reliable for comparison between participants.
For both parameters, the grid was linearly interpolated on both axes between the minimum and maximum values over 100 points. Then, the grid had a resolution of 4.6 revolutions for the magnitude of assistance and 1% gait cycle for timing. For clarity, the magnitude of assistance is reported on a normalized scale from 1 to 10.
Data analysis
During each preference trial, we recorded data from the prosthesis, including motor angle, motor velocity and ankle angle. From the IMU we extracted thigh angle and thigh velocity. We also recorded time series of all assistance settings (x, y) the participants experienced and the muscle activity signals from the 4 muscles specified in Sect. 2.1. The assistance settings are reported in magnitude, as explained in 2.4, but can also be represented using the resulting plantar flexion angle following push-off. This biomechanical measure ranged from 0 to 4.5
with a resolution of 0.5
.
We divided the outcomes into preference outcomes, exploration strategy outcomes and physiological outcomes. The preference outcomes include four main variables: the preferred magnitude of assistance (motor position), the preferred peak timing (% gait cycle), and the corresponding ankle angle and torque profile. The magnitude of assistance and peak timing represent the final settings selected by each participant. Since the goal was to provide assistance at push-off, we analyzed the different levels of plantar flexion torque and the corresponding ankle plantar flexion angles.
Then, for the exploration strategy outcomes, we have a number of settings explored before confirming their preference, the total exploration time (seconds) and number of strides per setting. Finally, for the physiological outcomes we have two key variables, the total muscle intensity and the synergies variability (explained in Sect. 2.6).
EMG data processing
Raw EMG recordings were band-pass filtered (10–450 Hz), full-wave rectified, and low-pass filtered (6 Hz) using a second-order Butterworth filter. For each participant and muscle, the resulting linear envelopes were normalized with respect to the maximum peak amplitude for that muscle. The latest was selected as the maximum value of a 50 ms moving average window applied to the muscle linear envelopes in each trial [29]. Next, we proceeded to process the EMG signals to calculate both muscle intensity and muscle synergies.
EMG intensity
The power spectral density (PSD) describes how the power of each EMG signal is distributed across frequencies. In this study, the PSD was estimated using Welch’s method (Hann window length 0.12s, overlap 0.50s). The PSD outputs the power per frequency (
).
To quantify each muscle’s power (or EMG intensity) [21], we computed the area under the PSD curve. This was done via numerical integration, as shown in Eq. 3, where f denotes the frequency vector.
The EMG intensity per muscle was computed individually, and the overall muscle intensity was obtained by summing the muscle intensity of the included muscles over the time period t.
![]() |
3 |
We computed the total muscle intensity for each participant and set of parameters. The TA was not included in the total muscle intensity since the gastrocnemius and RF have higher energy requirements [30]. The first and last two gait cycles were excluded from the analysis.
Muscle synergies
Muscle synergies were extracted using a non-negative matrix factorization (NNMF) from EMG signals. The EMG data was organized in a matrix format, with each row representing the time series of a specific muscle [31–33]. NNMF decomposes the muscle activity (
) into a linear combination of time-invariant synergy vectors (
) and activation coefficients (
). Each
displays the relative contribution of each muscle to synergy i, while the
represents how that synergy changes over time. For each set of parameters, two synergies were sufficient to explain over 90% of signal variance, in line with previous studies [34].
![]() |
4 |
For each set of parameters, we extracted muscle synergies from two time points: two strides from the beginning and two strides near the end of the condition. Strides were identified using the muscle activity from the GL. A stride was defined as starting at the maximum GL activity point (during stance) and ending after two stance phases [21]. Subsequently, muscle synergies were computed for the average EMG activity of each muscle over the selected strides. This approach ensured that participants had completed at least one full gait cycle, allowing for consistent comparison across conditions. However, if fewer than eight gait cycles were completed for a given set of parameters, it was not possible to extract synergies at two time points, and those sets were excluded from the analysis.
To measure muscle synergies variability, we compared the synergies extracted at the two distinct time points for each parameter set [21]. Higher variability suggests a lower level of coordination. To quantify variability, we computed the Euclidean distance between activation coefficient curves (
) at each time step. This metric is referred to as similarity between activation coefficients [21].
Results
The results section is organized into three parts. First, we present the exploratory strategy outcomes, which describe how much time each participant spent exploring different settings and how many settings were explored. Second, we report the preference outcomes, focusing on the sets of parameters selected as preferred across trials and participants. Third, we present the physiological outcomes, including muscle intensity and muscle synergy variability for the tested set of parameters.
Exploration strategy outcomes
Even though there were no restrictions on exploration time or number of settings allowed, the participants only needed between 110 and a maximum of 300 s to explore the parameters settings in each trial, testing no more than 14 parameters to find their preference (Table 1). For each parameter setting, participants 1 and 2 took more than 15 strides, around 25 s, before moving on to the next choice. Participant 3 sometimes had a lower number of steps per set of parameters, but this was sufficient to identify whether it was comfortable or not. Once the participant selected the preferred settings, they walked for an additional 20 s to confirm that these settings were indeed comfortable and preferable over all other options tested in the trial.
Table 1.
Summary of the exploration strategy outcomes per participant
| Participant | Number of settings explored | Total exploration time (sec) | Number of strides per setting | Time per setting (sec) |
|---|---|---|---|---|
| Participant 1 | ![]() |
![]() |
![]() |
![]() |
| Participant 2 | ![]() |
![]() |
![]() |
![]() |
| Participant 3 | ![]() |
![]() |
![]() |
![]() |
Most of the parameters tested were between 48 and 52% timing and the highest magnitudes of assistance were rarely tested. Participant 1, only in one trial selected the maximum magnitude which was reported as uncomfortable. As a result, high magnitudes were avoided in subsequent trials. Once the participant recognized the direction in which the amplitude increased on the grid, the participant chose not to explore that area. This exploration strategy reduced the exploration area and allowed the participant to exploit more comfortable settings faster. In contrast, participants 2 and 3 explored higher assistance levels. Although their subjective feedback on these settings was generally positive, the sensations were unfamiliar and did not align with their expectations, and as a result, these settings were not selected as preferred.
Preference outcomes
As shown in Fig. 3A, the preferred set of parameters varied across trials for participant 1 and 2, as reflected in the larger variability in both timing and magnitude of assistance. In most trials, these participants selected lower assistance magnitudes, but in one trial each, a higher magnitude was preferred, contributing to the variability shown in Fig. 3A. In contrast, participant 3 consistently selected similar parameter settings across trials, with no variation in the magnitude of assistance.
Fig. 3.
A Mean preferred assistance magnitude (also represented as ankle plantar flexion angle) and timing across the four trials for each participant. Error bars represent standard deviation. B Preferred torque profiles corresponding to the selected assistance parameters for each participant
The mean preferred torque profiles across trials and for all participants are presented in Fig. 3B. The torque was calculated based on the commanded motor position and the measured ankle angle using a lookup table, as described in Sect. 2.3.2. This figure highlights the importance of personalized control, as participants showed distinct trends in their preferred assistance torque profiles. Although variability across trials leads to some overlap, the individual mean profiles suggest different preferences in timing and magnitude of assistance.
Moreover, the low torque values observed for participant 1 may reflect limited engagement with the motor and passive components of the device, resulting in reduced energy delivery (see Additional file 1: Fig. S3-S4). This is aligned with the torque–angle curve, which indicates reduced torque across the full range of motion (Additional file 1: Fig. S5). In contrast, participants 2 and 3 showed higher torque profiles, particularly participant 3 shows consistent torque profiles, which align with expected behavior given their selected assistance magnitudes. It is also important to mention that although participant 3 consistently selected the same magnitude of assistance (Fig. 3A), variability in the torque profile remains due to differences in ankle kinematics across steps and trials.
Physiological outcomes
Figure 4 shows total muscle intensity per participant in a heatmap, where the y-axis represents assistance magnitude and the x-axis represents assistance timing. The colors in the heatmap indicate the total muscle intensity. Since participants did not explore the full grid it is difficult to identify a precise area of preference for timing and magnitude but still some trends are observed.
Fig. 4.
Heatmaps showing the total muscle intensity for all the settings tested across trials for each participant. Blue indicates low muscle intensity and yellow indicates high muscle intensity. The y-axis shows the assistance magnitude, and the x-axis shows the assistance timing. Red boxes highlight the preferred settings selected across the trials, and orange boxes indicate settings that were chosen twice
For all participants, lower muscle intensity values tend to cluster around assistance timings between 48 and 50% of the gait cycle. Regarding the magnitude of assistance, the results varied among the participants. For example, participant 2 showed lower muscle intensity with lower magnitudes, although other parameter combinations also produced low muscle intensity. These settings corresponded to higher dorsiflexion and low plantar flexion angle after push-off. In contrast, participants 1 and 3 had lower muscle intensity around the middle of the grid (i.e., magnitude of 6), which was associated with lower dorsiflexion values and higher plantar flexion angles following push-off.
Interestingly, participant 1 reported that the maximum assistance magnitude felt very uncomfortable, which was reflected in a high total muscle intensity value (1.3
).
The preferred settings did not always match the lowest total muscle intensity in all trials. However, when analyzing individual muscles, each participant’s preferred settings were associated with reduced activity in specific muscles. This suggests that each participant may prioritize reducing effort in particular muscle groups based on personal comfort or perception. For example, participant 1 preferred settings that were associated with reduced intensity in the GL, GM and RF, but not in the TA (see Additional file 1: Fig. S6-S9).
Regarding muscle synergies variability, no clear trend was observed in the heatmap across trials (Fig. 5). However, within each participant, the preferred settings consistently showed lower variability. These findings indicate that, on a trial-by-trial basis, the preferred settings are linked to lower synergy variability, which reflects better stride coordination.
Fig. 5.
Heatmaps showing the muscle synergies variability for all the settings tested across trials for each participant. Blue indicates low variability in muscle synergies, while yellow indicates high variability. The y-axis shows the assistance magnitude, and the x-axis shows the assistance timing. Red boxes highlight the preferred settings selected across the trials, and orange boxes indicate settings that were chosen twice
Discussion
This study explored a preference-based optimization of an ankle prosthesis during the push-off phase of walking. Using a self-exploration approach [12], the participants were allowed to explore a variety of settings and select the set of settings they found most comfortable. Furthermore, this research investigated how preference is related to muscle activity intensity and muscle synergy variability using some of the main muscles involved during walking.
Previous research on optimizing an ankle prosthesis using user preference has also used a self-exploration strategy paradigm, where users adjust the device’s behavior and select their preferred settings according to their own internal strategy [13–15]. This method allows exploration of the parameter space (grid). The participants then intuitively focus on the area that feels more comfortable. This strategy is particularly effective when tuning a low-dimensional set of parameters, like in this study.
While optimizing for user preference shows promising results, as it can reflect the overall user experience, a key challenge is to identify what drives preference [12]. Clites et al. [15] evaluated the kinematics of individuals with transtibial amputations walking with their preferred stiffness level, as well as with higher and lower stiffness values. They reported that lower gait asymmetry explained the participant’s preference, although there was no improvement in metabolic cost. Thus, is still unclear which physiological factors are considered in preference, which motivated the exploration of muscle activity.
From the results, we were particularly interested in the relationship between user preference, muscle synergy variability and total muscle intensity. The participants selected their preferred settings when their motor coordination was higher, as indicated by reduced muscle synergy variability. At the same time, the muscle intensity from most of the measured muscles decreased. However, these reductions were not consistent across participants. Each participant showed reduced muscle activity in different muscles, indicating that each individual may prioritized different groups of muscles when selecting their preference [35]. Although the TA did not show reduced EMG intensity in the preferred settings for some participants, this muscle likely played a minor role in preference selection, given the higher metabolic demands of the gastrocnemius and rectus femoris [30]. Based on these insights and previous research, it would be worth exploring gait kinematics and investigating if there is a relationship between gait symmetry and muscle synergies variability in individuals with lower-limb amputation [36–38].
Despite other researchers not finding a relation between preference and a decrease in metabolic cost [15], the present findings suggest that the participant’s preference was influenced by reductions in total muscle intensity and comfort. Both of these factors have been associated with energy cost in previous research [39]. Muscle activity captures important factors such as effort, motor coordination and overall efficiency. As shown in the results, muscle synergy variability also appears to be a promising predictor of user preference, although this remains an emerging trend that requires further investigation.
Limitations
Most individuals typically use a passive or quasi-passive prostheses, and adapting to a new active prosthesis requires time. We recognize this as one of the main limitations of the study, since we only had one familiarization session. Although participants reported feeling confident and overall comfortable with the prosthesis, their long-term use of passive devices likely influenced their preferences. Specifically, participants tended to prefer lower levels of push-off, possibly because higher plantar flexion torques felt unfamiliar or unnatural [40]. For example, one participant selected low assistance magnitudes in most trials closely resembling their usual walking experience. As a result, the exploration time was relatively short per trial, and a narrower range of tested parameters. With additional familiarization sessions, participants may become more accustomed to the dynamic characteristics of the active prosthesis and might be more open to exploring higher assistance magnitudes, as well as higher plantar flexion during push-off. Another study reported a mild positive relationship between preferred magnitude and walking speed [14]. Since we used a fixed speed across all trials, it is possible that allowing different walking speeds could also lead to a preference for higher assistance magnitudes.
Another limitation of this study is its small sample size, which limits the findings to descriptive observations rather than allowing statistical analysis. Increasing the sample size in future research would be valuable in determining whether the observations are statistically significant.
Furthermore, torque values cannot be consistently compared across participants, as is common in the literature. This is because torque is estimated based on motor position and ankle angle, the latest being influenced by the user’s movement. Since ankle kinematics vary between participants and, sometimes, even across steps, the resulting torque estimates are not consistent. This added complexity made the presentation of the results more difficult. As a result, improving the control strategy should be a priority in future work.
Despite the limitations, we demonstrate that the current hardware can provide different assistance settings, and that participants are able perceive these differences and select a preferred option. In addition, the results show insights into how user preference is related to muscle effort and motor coordination. Future work will focus on expanding the sample size and validating this approach, with the aim of developing a data-driven model capable of estimating user preference based on muscle activity. This approach could even be combined with biomechanical data, building on prior positive results in estimating preference from biomechanics [20]. A robust user preference strategy has the potential to inform the design and prescription of clinical assistive devices that increase user satisfaction [15].
In addition, this approach could be tested and extended to other daily tasks relevant to the CYBATHLON competition.
Conclusion
This study shows that allowing individuals with lower-limb amputation to adjust ankle assistance during walking according to their preference and comfort can show valuable insights into the relationship between user preference and physiological factors. Our findings, show that user-preferred assistance settings were generally associated with lower muscle intensity across most measured muscles, suggesting that users select assistance levels that reduce muscular effort. This highlights the role of muscle activity as a key physiological factor influencing user preference for prosthetic push-off assistance. Integrating user preference optimization into assistive robotic devices has great potential to enhance device personalization and promote better mobility outcomes for users.
This work opens several opportunities for future research. First, future studies should explore the generalizability of these methods in larger participant populations. In addition, a promising direction is to include learning-based optimization strategies that simultaneously consider both muscle activity and user preference during the adaptation process. These approaches could improve the convergence time, and enhance the user experience while reducing muscle effort.
Supplementary information
Acknowledgements
The authors thank the Amptraide organization for helping us to find participants and their interest in developing this work. Also, special thanks goes to Ellen Roels for her help and support in developing the interactive interface used in this study.
Author contributions
M.D., L.F., P.B., K.P., and T.V. conceptualized the study. M.D., S.N., and L.F. designed the experiments and controllers. M.D., S.N., and L.F. conducted the experiments. M.D. analyzed muscle activation and synergies data. S.N. preprocessed the data collected from the prosthesis. M.D., S.N, and L.F. interpreted the data. M.D., P.B, T.V and K.P. prepared the initial manuscript. All authors reviewed the manuscript.
Funding
This work was supported by BOSA FOD AI-Driven Wearable Robotics for Healthcare (Aidwear) and the Strategic Research Program Exercise and the Brain in Health and Disease: The Added Value of Human-Centered Robotics, Vrije Universiteit Brussel, Belgium.
Data availability
The datasets collected and/or analyzed in the current study are available from the corresponding author on reasonable request.
Declarations
Ethical approval and consent to participate
This study followed the standards of the Declaration of Helsinki, and the local medical ethical commission (Vrije Universiteit Brussel and UZ Brussel) granted ethical approval (B.U.N. 143201941703). Participants provided written informed consent prior to the study.
Consent for publication
The participants signed consent for publication.
Competing interests
The authors declare no Conflict of interest.
Footnotes
Kevin De Pauw is the primary corresponding author of the article.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Contributor Information
María Alejandra Díaz, Email: ma.diaz@vub.be.
Kevin De Pauw, Email: Kevin.De.Pauw@vub.be.
References
- 1.Leal-Junior A, Frizera-Neto A. Chapter 1 - introduction and overview of wearable technologies. In: Leal-Junior A, Frizera-Neto A (eds) Optical fiber sensors for the next generation of rehabilitation robotics, pp. 3–26. Academic Press, Cambridge, MA, USA (2022). https://doi.org/10.1016/B978-0-32-385952-3.00009-3https://www.sciencedirect.com/science/article/pii/B9780323859523000093.
- 2.Price MA, Beckerle P, Sup FC. Design optimization in lower limb prostheses: a review. IEEE Trans Neural Syst Rehabil Eng. 2019;27(8):1574–88. [DOI] [PubMed] [Google Scholar]
- 3.Díaz MA, et al. Human-in-the-loop optimization of wearable robotic devices to improve human-robot interaction: a systematic review. Cybern IEEE Trans. 2022;53(12):7483–96. [DOI] [PubMed] [Google Scholar]
- 4.Zhang J, Fiers P, Witte KA, Jackson RW, Poggensee KL, Atkeson CG, et al. Human-in-the-loop optimization of exoskeleton assistance during walking. Science. 2017;356(6344):1280–4. [DOI] [PubMed] [Google Scholar]
- 5.Song S, Collins SH. Optimizing exoskeleton assistance for faster self-selected walking. IEEE Trans Neural Syst Rehabil Eng. 2021;29:786–95. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Jackson RW, Collins SH. Heuristic-based ankle exoskeleton control for co-adaptive assistance of human locomotion. IEEE Trans Neural Syst Rehabil Eng. 2019;27(10):2059–69. [DOI] [PubMed] [Google Scholar]
- 7.Welker CG, Voloshina AS, Chiu VL, Collins SH. Shortcomings of human-in-the-loop optimization of an ankle-foot prosthesis emulator: a case series. R Soc Open Sci. 2021;8:202020. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Kim M, et al. Bayesian optimization of soft exosuits using a metabolic estimator stopping process. In: IEEE Int Conf Robot Autom, 2019:9173–9179 . 10.1109/ICRA.2019.8793817.
- 9.Tucker M, et al. Human preference-based learning for high-dimensional optimization of exoskeleton walking gaits. In: IEEE/RSJ Int Conf Intell Robots Syst, 2020:3423–3430 10.1109/IROS45743.2020.9341416.
- 10.Tucker M, et al. Preference-based learning for exoskeleton gait optimization. In: IEEE Int Conf Robot Autom, 2020:2351–2357 10.1109/ICRA40945.2020.9196661.
- 11.Wen TC, Jacobson M, Zhou X, Chung HJ, Kim M. The personalization of stiffness for an ankle-foot prosthesis emulator using human-in-the-loop optimization. In: IEEE/RSJ Int Conf Intell Robots Syst, 2020:3431–3436. 10.1109/IROS45743.2020.9341101.
- 12.Ingraham KA, Tucker M, Ames AD, Rouse EJ, Shepherd MK. Leveraging user preference in the design and evaluation of lower-limb exoskeletons and prostheses. Curr Opin Biomed Eng. 2023;28:100487. [Google Scholar]
- 13.Shepherd MK, Azocar AF, Major MJ, Rouse EJ. Amputee perception of prosthetic ankle stiffness during locomotion. J Neuroeng Rehabil. 2018;15(1):1–10. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Ingraham K, Remy C, Rouse E. The role of user preference in the customized control of robotic exoskeletons. Sci Robot. 2022;7(64):eabj3487. [DOI] [PubMed] [Google Scholar]
- 15.Clites TR, Shepherd MK, Ingraham KA, Wontorcik L, Rouse EJ. Understanding patient preference in prosthetic ankle stiffness. J Neuroeng Rehabil. 2021;18(1):1–16. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Lee UH, Shetty VS, Franks PW, Tan J, Evangelopoulos G, Ha S, et al. User preference optimization for control of ankle exoskeletons using sample efficient active learning. Sci Robot. 2023;8(83):3705. [DOI] [PubMed] [Google Scholar]
- 17.Hamaya M, et al. Exploiting human and robot muscle synergies for human-in-the-loop optimization of emg-based assistive strategies. In: IEEE Int Conf Robot Autom, 2019:549–555. 10.1109/ICRA.2019.8794082.
- 18.Alili A, Nalam V, Li M, Liu M, Feng J, Si J, et al. A novel framework to facilitate user preferred tuning for a robotic knee prosthesis. IEEE Trans Neural Syst Rehabil Eng. 2023;31:895–903. [DOI] [PubMed] [Google Scholar]
- 19.Caputo JM, Dvorak E, Shipley K, Miknevich MA, Adamczyk PG, Collins SH. Robotic emulation of candidate prosthetic foot designs may enable efficient, evidence-based, and individualized prescriptions. JPO J Prosthet Orthot. 2022;34(4):202–12. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Shetty VS, Lee UH, Ingraham KA, Rouse EJ. A data driven approach for predicting preferred ankle stiffness of a quasi-passive prosthesis. IEEE Robot Autom Lett. 2022;7(2):3467–74. [Google Scholar]
- 21.Díaz MA, et al. Human-in-the-loop optimization of wearable device parameters using an emg-based objective function. Wearable Technol. 2024;5:15. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Hameed I, Tuan P, Camana M, Koo I. Wearable sensory apparatus for real-time feedback in wearable robotics. Electronics. 2021;10(509):1–16. [Google Scholar]
- 23.MacLean MK, Ferris DP. Human muscle activity and lower limb biomechanics of overground walking at varying levels of simulated reduced gravity and gait speeds. PLoS ONE. 2021;16(7):0253467. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Di Nardo F, Ghetti G, Fioretti S. Assessment of the activation modalities of gastrocnemius lateralis and tibialis anterior during gait: a statistical analysis. J Electromyogr Kinesiol. 2013;23(6):1428–33. [DOI] [PubMed] [Google Scholar]
- 25.Flynn L, et al. The challenges and achievements of experimental implementation of an active transfemoral prosthesis based on biological quasi-stiffness: the cyberlegs beta-prosthesis. Front Neurorobot. 2018;12:80. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Ingraham KA, Fey NP, Simon AM, Hargrove LJ. Assessing the relative contributions of active ankle and knee assistance to the walking mechanics of transfemoral amputees using a powered prosthesis. PLoS ONE. 2016;11(1):0147661. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Quintero D, Villarreal DJ, Lambert DJ, Kapp S, Gregg RD. Continuous-phase control of a powered knee–ankle prosthesis: amputee experiments across speeds and inclines. IEEE Trans Rob. 2018;34(3):686–701. 10.1109/TRO.2018.2794536. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Malcolm P, Quesada RE, Caputo JM, Collins SH. The influence of push-off timing in a robotic ankle-foot prosthesis on the energetics and mechanics of walking. J Neuroeng Rehabil. 2015;12:1–15. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Úbeda A, Azorín JM, Farina D, Sartori M. Estimation of neuromuscular primitives from eeg slow cortical potentials in incomplete spinal cord injury individuals for a new class of brain-machine interfaces. Front Comput Neurosci. 2018;12:3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Umberger BR, Gerritsen KG, Martin PE. A model of human muscle energy expenditure. Comput Methods Biomech Biomed Engin. 2003;6(2):99–111. [DOI] [PubMed] [Google Scholar]
- 31.d’Avella A, Saltiel P, Bizzi E. Combinations of muscle synergies in the construction of a natural motor behavior. Nat Neurosci. 2003;6(3):300–8. [DOI] [PubMed] [Google Scholar]
- 32.Chvatal SA, Ting LH. Common muscle synergies for balance and walking. Front Comput Neurosci. 2013;7:48. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Cheung VC, et al. Plasticity of muscle synergies through fractionation and merging during development and training of human runners. Nat Commun. 2020;11(1):43–56. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34.Zhao K, et al. Intra-subject and inter-subject movement variability quantified with muscle synergies in upper-limb reaching movements. Biomimetics. 2021;6(4):63. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35.Ting LH, McKay JL. Neuromechanics of muscle synergies for posture and movement. Curr Opin Neurobiol. 2007;17(6):622–8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36.Feng L, Yu L, Lyu H, Yang C, Liu X, Zhou C, et al. Synergy in motion: exploring the similarity and variability of muscle synergy patterns in healthy individuals. Hum Mov Sci. 2024;98:103300. [DOI] [PubMed] [Google Scholar]
- 37.Safavynia S, Torres-Oviedo G, Ting L. Muscle synergies: implications for clinical evaluation and rehabilitation of movement. Top Spinal Cord Injury Rehabil. 2011;17(1):16–24. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38.Winter DA. Biomechanics of normal and pathological gait: implications for understanding human locomotor control. J Mot Behav. 1989;21(4):337–55. [DOI] [PubMed] [Google Scholar]
- 39.Blake OM, Wakeling JM. Estimating changes in metabolic power from emg. Springerplus. 2013;2(1):1–7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 40.Beckerle P, et al. A human-robot interaction perspective on assistive and rehabilitation robotics. Front Neurorobot. 2017;11:24. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The datasets collected and/or analyzed in the current study are available from the corresponding author on reasonable request.




















