Structured modulations in high-density EMG patterns from a single muscle enable simultaneous control of natural and extra degrees of freedom Authors: Julien Rossato, Elodie Hinnekens, Daniele Borzelli, Sergio Gurgone, Denise Jennifer Berger, Andrea d’Avella Published: 2026 ABSTRACT AbstractBackgroundIdentifying and exploiting independent neural inputs within a single muscle is a key step toward improving neuromuscular interfaces for motor augmentation and rehabilitation. Recent evidence suggests that motor neurons within the same pool can receive multiple, independent inputs, allowing flexible modulation of muscle activity. Harnessing this neural flexibility could enable the control of supplementary degrees of freedom (SuDoF) without interfering with the performance of natural motor tasks. However, current methods rely on motor unit decomposition, which is technically challenging and impractical for daily-life applications. This study introduces a decomposition-free approach to detect and quantify distinct neural inputs directly from high-density surface electromyography (HDsEMG) and assesses the feasibility of extracting from the modulation of these neural inputs a signal for simultaneous control of natural and SuDoF.MethodsFifteen participants performed reaching and orienting tasks in a virtual environment moving a cursor by generating isometric force at the hand (force control) and rotating the cursor by modulating the pattern of HDsEMG signals recorded using a 64-channel grid from the biceps brachii (SuDoF control). HDsEMG signals were represented as vectors in a multidimensional EMG space, and its deviation from a reference vector associated with pure force control was used to control the cursor orientation. Participants reached eight planar targets with a circular cursor using only force control and reached and matched the orientation of eight elliptical targets with an elliptical cursor using force and SuDoF control. Performance across conditions was analyzed using linear mixed-effects models.ResultsParticipants successfully reached targets in 89 ± 10% of force and SuDoF control trials, with cursor rotations exceeding those observed in force-only trials in 84 ± 31% of cases. Participants primarily modulated contraction amplitude but maintained distinct directional shifts in HDsEMG patterns, compatible with the modulation of different neural inputs within the motor neuron pool. However, these modulations may also partly reflect changes in overall muscle activation and recruitment.ConclusionsThis study presents a decomposition-free method to detect structured modulations in HDsEMG patterns within a single muscle. The observed modulations are compatible with the presence of distinct neural inputs and enabled concurrent control of natural and SuDoF, advancing neuromuscular interface design for assistive and rehabilitative technologies.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12984-026-02029-z.Keywords: High-density electromyography, Muscular null space, Neural inputs detection, Human augmentation, Myoelectric control, Reaching Background How to increase the number of degrees-of-freedom (DoF) that an individual can control is a growing research topic due to the possibilities it opens up for controlling supernumerary robotic limbs [1, 2]. Robotic augmentation has applications both for individuals with disabilities, for compensating impairments, and for healthy individuals, to enhance their abilities [3, 4]. While individuals born with extra DoFs, like those with six-fingered hands, exhibit superior manipulation abilities [5], how well natural and artificial DoF can be coordinated remains an open question. Control of artificial, supplementary DoF (SuDoF) must not interfere with the activity of natural DoF [6]. One possibility is to leverage the redundancy of the musculoskeletal and neural systems [7]. Recent studies have demonstrated that it is possible to exploit the muscular null space to control SuDoF [8, 9]. Muscular null space refers to those muscle activation patterns that do not affect the movement or force output required to perform a specific task. For example, activating a muscle that is not involved in the task lies within the muscular null space for that task. A more intriguing example is the co-contraction of a group of muscles involved in the task, where their forces counteract each other, resulting in no net contribution to the task’s goal. Furthermore, some studies have shown that it is possible to control SuDoF and natural DoF simultaneously [10]. While muscular null space may be a promising approach to control SuDoF, current methods require recording from many muscles, which can be challenging in real-world applications. In contrast, a method to extract muscular null space from the electromyographic (EMG) signals recorded from a single muscle would simplify the design of a wearable interface. In the case of a muscle directly involved in the task, it should be possible to extract a muscular null space signal if this muscle is controlled by at least two independent neural inputs. Modulation of two neural inputs may be achieved while keeping the same force output, thus corresponding to muscular null space modulation. Recent studies have shown that multiple neural inputs drive the motor neurons within the same pool [11–15]. These findings have led to the notion of motor neuron synergy [16], i.e., a group of motor neurons within the pools of different muscles receiving the same neural input and thus representing a functional module used by the CNS to reduce the dimensionality of the control. Modulation of each independent neural input innervating a specific group of motor neurons within the pool of a single muscle could then be exploited to control SuDoF independently of force output, i.e. as a control signal in the muscular null space. Specifically, Formento et al. [17] demonstrated that it is possible to control up to three motor units within the biceps brachii independently. This control over the muscle’s dimensionality enhances the potential for null space modulation and reduces the number of muscles that need to be recorded. The identification of independent neuronal inputs can be achieved through temporal or frequency-based correlation [14, 18–20] or by using dimensionality reduction techniques [21, 22] on motor unit discharge activity. To do this, it is first necessary to obtain motor unit spike trains, which can be obtained from the decomposition of high-density surface EMG (HDsEMG) or high-density intramuscular EMG [23–25]. However, identifying motor units using HDsEMG is challenging and often unsuccessful. Previous studies often favored recruiting athletic individuals to increase the number of motor units that could be tracked, and signals from female participants typically enabled the identification of only half as many motor units [26, 27]. Identifying motor units requires that each unit exhibits a unique action potential profile or shape. However, the location of the muscle fibers relative to the electrodes and the conduction velocity of action potentials impact the low-pass filtering effect of the tissue on recorded data and ultimately affect motor unit shape [28]. A method that would work directly at the level of global EMG pattern would overcome these limitations and simplify practical applications. If independent neural inputs are distributed differently to the motor neurons of a single muscle, each input could also produce a distinct temporal and spatial EMG profile through the summation of the unique shapes of the motor unit action potentials. Therefore, it would be theoretically possible to track independent neural inputs reaching motor neuron synergistic groups and extract the muscular null space from a single muscle without identifying motor units or using EMG decomposition techniques. By assuming that (i) multiple neural inputs drive different subsets of motor neurons within the pool of a single muscle [16], and (ii) the HDsEMG signals recorded by the electrodes contain unique pattern signatures for each neural input, we propose a new method to identify muscular null space modulation (Fig. 1). The existence of unique pattern signatures implies that, when a single neural input is modulated, the vector comprising all the EMG values should align with a distinct direction in the EMG space. Any deviation from this direction could be attributed either to noise or to the activation of another neural input with a different pattern signature, and thus a different direction. This deviation can be quantified using the distance or the angle between the two EMG signal vectors (Fig. 1D). However, it remains unclear whether such shifts caused by new neural inputs are sufficiently distinct to be differentiated from noise-induced variations. Identification of muscular null space modulation. A Conceptual framework behind the new method to identify muscular null space modulation. Two neural inputs reach two subsets of motor neurons pool innervating a muscle. Each neural input produces a unique pattern recorded by the HDsEMG. and are vectors comprising all the EMG values. B Theoretical representations of the unique patterns recorded by the HDsEMG. Raw EMG of each channel are represented at the top. Average EMG amplitudes of each channel are represented at the bottom. C Representation of activation pattern leading to the same force production by the muscle from the conceptual framework A. All these activation patterns belong to the muscular null space. D Geometric representation of the Distance value D and the angle between This study examines the feasibility of detecting different neural inputs within the biceps brachii using HDsEMG, without applying EMG signal decomposition techniques. It also explores the possibility of using HDsEMG signal to control SuDoF while concurrently controlling natural DoF. We used an experimental paradigm that requires simultaneous modulation of hand force and muscular null space activation, and we introduced a new method based on the direction and amplitude of the vector containing the instantaneous HDsEMG signal in a vector space with as many dimensions as EMG channels. Sixteen participants performed isometric force production tasks to control a cursor in a virtual environment. In a force control task, the displacement of a circular cursor was controlled by the force. In a force and SuDoF control task, the rotation of an elliptical cursor was controlled with a muscular null space modulation while its position was controlled by force. This experimental protocol explores the feasibility of the simultaneous control of natural DoF and one SuDoF using only the HDsEMG signals from the biceps brachii. Methods Participants Sixteen volunteers (mean ± SD; age: 32.8 ± 11.2 years, age range: 21–64, 6 females, 3 left-handed) participated in the experiments after giving written informed consent. All procedures were conducted in accordance with the principles embodied in the Declaration of Helsinki, complied with national regulations, and were approved by the Lazio territorial ethics committee Area 5 (N.82/SL/23). All participants had normal or corrected-to-normal vision and did not report any known neurological disorder or upper dominant limb injury. Experimental setup Participants sat on a chair with their dominant arm kept along the trunk and the forearm resting on a table (Fig. 2A). The chair was positioned so that participants could hold a handle connected to a 6-axis force/torque transducer (Nano25 Transducer, ATI Industrial Automation, Apex, USA), lean against the back of the chair and have a 90° angle at the elbow (0° being the full extension of the arm). Depending on the participant’s height, the dominant arm was positioned such that it was abducted by around 60° with respect to the trunk in the frontal plane. Wooden planks were positioned under the participant’s forearm to adjust its height so that the axis of the forearm and the y-axis of the transducer were aligned. A computer monitor was positioned in front of the participant and used to display a cursor and simple virtual scenes for various experimental tasks. These tasks were implemented in MATLAB on a personal computer (PC) using forces and torques data recorded by the transducer and EMG data recorded from HDsEMGs (see section ‘Surface electromyography’ below). Forces and torques data were digitized at 2000 Hz using a data acquisition card (NI PCI-6229, National Instruments Corporation, Austin, USA) in the PC. EMG data were collected from the EMG system through a USB port of the PC. Experimental setup and protocol. A Experimental setup: participants sat on a chair with their dominant arm resting on a table close to their side. The chair was positioned so that participants could hold a handle connected to a 6-axis force transducer. A computer monitor was positioned in front of the participants and displayed a cursor providing feedback to perform various tasks. HDsEMG signals were recorded from the biceps brachii. A reference electrode was positioned over the medial epicondyle of the dominant arm’s elbow. A strap electrode dampened with water was placed around the contralateral wrist (ground electrode). B All trials from the Force control task are shown for a representative participant (Participant 4). The trials are color-coded according to the target position. Cursor trajectories are shown in separate plots for target position (red crosses) at different distances. C Schematic of the numeric weights applied to the monopolar signals of adjacent HDsEMG electrodes to compute the inverse binomial spatial filter. This filter consisted of the summation of the signal from one electrode with the signals from its 8 surrounding electrodes while multiplying each signal by the weights shown in the diagram (-12 for the central electrode, 2 for electrodes in the same row or column, 1 for the electrodes along the diagonal). D Target positions during Force control task (left) and Force and SuDoF control with target at 0° (center) or with target at 45° (right). Black circles or ellipsoids represent the cursor at rest position. E Schematic of the experimental protocol and extracted parameters from each task (MVC = maximum voluntary contraction) Experimental procedure The entire experimental procedure is depicted in Fig. 2E. The experimental session began with a standardized warm-up, which consisted of pulling the handle for 5 s at 50%, 60%, 70%, 80% and 90% of the participants’ subjective maximal force with a rest period in-between. Then, after 1-minute rest period, participants performed two maximal voluntary contractions (MVC) by pulling the handle as hard as they could for 3 s, with 60 s of rest in between. Following this, participants generated isometric pronation torques at 50% of the participants’ subjective maximal torque for 3 s, with 60 s of rest in between. Peak MVC force (MVF) and torque were considered as the maximal average values obtained within a moving window of 250 ms. After 2 min of rest, participants performed a Force control task in which they were instructed to quickly and accurately move a cursor from a resting position to a target by applying isometric forces on the handle and then hold the cursor at the target position. The cursor moved on the screen according to the horizontal forces applied on the handle: pushing the handle forward (i.e., along the anteroposterior body axis) moved the cursor along the vertical direction (y-axis of the computer monitor) and pushing it to the right (i.e., along the mediolateral body axis) moved the cursor along the horizontal direction to the right (x-axis of the computer monitor). In total, 72 targets appeared. They were randomly located in one of 24 possible spatial positions surrounding the resting position on the monitor, corresponding to the combination of 8 different directions (uniformly distributed around a circle at 45 degrees intervals, starting from the direction of the x-axis) and 3 different distances (corresponding to 10%, 20%, and 30% of MVF). For each trial, participants had 4 s to reach the target and had to maintain the cursor within the target for 1 s, with a tolerance of 10% of the target distance. After successfully completing the trial, or after 4 s if unable to reach the target, participants rested for 2 s. Next, participants performed a Holding task in which they were instructed to quickly and accurately move a cursor from a resting position to a target by applying isometric forces or torques on the handle, then hold the cursor for 20 s. The task involved seven targets. Four targets were identical to the diagonal targets at 20% of MVF from the Force control task. For two additional targets, participants had to produce both pronation and supination torques at 40% of their subjective maximal torque (based on 50% of the participants’ subjective maximal torque). The last target required participants to maintain a co-contraction of the biceps and triceps without producing any force or torque, matching the average muscle activation of the previous six targets (average of the rectified signal across all HDsEMG channels during the six previous contractions after low-pass filtering with a cutoff frequency of 5 Hz). Data from the Holding task were used to decode motor units’ discharge patterns (see below the “Identification of motor unit discharge patterns” section for details). Participants were then tested on their ability to control a SuDoF by altering the pattern of activation of the MUs of their biceps muscle, and consequently the pattern of EMG signals recorded on the HDsEMG electrode array, with respect to the pattern recorded during the Force control task. To do so, we computed a ‘Distance’ value from the HDsEMG signals recorded from the biceps (see below the “Extract null space modulation from HDsEMG” section for details). First, participants performed a SuDoF control training task in which they were trained to control this Distance value without simultaneously engaging their natural DoF involved in the force control task, i.e., without generating any end-point force. During training, participants received visual feedback on their force and torque production, as well as the Distance value. In the 2D-space of the computer monitor, a cursor moved along the x-axis proportionally to the magnitude of the force applied to the handle, and participants were instructed to keep the force under 5% of their MVF. The cursor’s color indicated torque production, transitioning from blue to purple as torque increased and turning red when the torque exceeded 10% of the participant’s subjective maximal torque (based on 50% of the participants’ subjective maximal torque). Participants were instructed to keep the torque below this 10% threshold. The cursor moved along the y-axis in response to changes in the Distance value, and participants were instructed to attempt to increase this value for 5s. The training was divided into blocks of five trials. Participants performed between one and 11 blocks, depending on how long they took to successfully complete the trials, while ensuring that the training duration did not reduce the time available for subsequent tasks. Success in a trial was defined as maintaining force and torque below the required thresholds while achieving a Distance value greater than the 95th percentile of the Distance values obtained during the Force control task. Additionally, if participants easily controlled the cursor along the y-axis and completed the task quickly, they were orally encouraged to reach a Distance value 3 to 4 times higher than the displayed threshold. After training SuDoF control, participants were tested on their ability to simultaneously control their natural DoF and the SuDoF. They performed a Force and SuDoF control task, in which they were instructed to move an elliptical cursor from a resting position to a target position (similar to the Force control task) and simultaneously adjust the orientation of this elliptical cursor to match the target orientation, as in a previous study using multi-muscle null space control [10]. Once the cursor matched the target position and orientation, participants had to hold the cursor at the target for 1 s, with a 10% tolerance for both position and orientation. To adjust the cursor’s orientation, participants increased the Distance value, where a 10° rotation corresponded to the 95th percentile of the Distance values obtained during the Force control task. The Force and SuDoF control task was subdivided into five blocks of 16 trials. In each block, targets appeared randomly in 8 different directions, at a distance corresponding to a force magnitude of 15% MVF, and with an orientation requiring either a 45° rotation or no rotation (0°). For each trial, participants had 5 s to reach the target position and orientation and to maintain the cursor on the target for 1 s. After participants completed the trial, or after 5 s if they did not, they rested for 2 s. Surface electromyography High-density surface EMG (HDsEMG) signals were recorded from the biceps brachii. A two-dimensional adhesive grid of 64 electrodes (8 × 8 electrodes, Ag/AgCl electrodes, inter-electrodes distance: 8.75 mm; Textile HD-EMG Grids, TMSi, Oldenzaal, The Netherlands) was placed over the muscle belly with the boundary between the short and the long head placed in the middle of the grid. Before the grid placement, the skin was cleaned with an abrasive gel (Nuprep, Weaver and Company, USA). The grid was held on the skin using a disposable adhesive layer. The cavities within the adhesive layer were filled with conductive gel (ECI Electro-gel, Electro-Cap International Inc., Eaton, USA) to facilitate the skin-electrode contact. A 10-cm-wide elastic band was placed over the grid to ensure good contact between the electrodes and the skin throughout the experiment. A reference electrode was positioned over the medial epicondyle of the dominant arm’s elbow. A ground electrode dampened with water was placed around the contralateral wrist. To improve spatial resolution, we applied an inverse binomial spatial filter to the HDsEMG signals [29, 30]. This binomial spatial filter consisted of the summation of a signal channel with its 8 surrounding neighborhoods while multiplying each signal by the appropriate weight shown in Fig. 2C. Before applying this spatial filter, the HDsEMG signals were high-pass filtered at 10 Hz using a Butterworth filter (3rd order) with zero-lag (‘butter’ and ‘filtfilt’ Matlab functions). After applying the binomial spatial filter, the signals were subsequently rectified and filtered again using a 5 Hz low-pass filter (2nd order). Channels located at the grid’s border, where the inverse binomial spatial filter could not be computed, were not considered. Channels with noisy signals or with neighboring noisy channels, were excluded. On average, 3.6 ± 5.9 channels were discarded (ranging from 0 to 21), resulting in an EMG vector containing 29.6 ± 9.3 values after spatial filtering (ranging from 10 to 36). Finally, the signals were normalized to their maximum amplitude within a 250-ms moving window during the Force Control task, and an offset, based on a 1-s resting signal recording, was subtracted. The EMG signals were recorded in monopolar mode from the HDsEMG on the biceps brachii. The signals were digitized at a sampling rate of 2000 Hz (SAGA 64+, EMG amplifier, TMSi, Oldenzaal, The Netherlands). Synchronization between the EMG signals and the force and torque signals was achieved by sending a TTL signal to both the force/torque acquisition card and the EMG recording system. Extracting null space modulation from HDsEMG The rationale for extracting null space modulation from HDsEMG is grounded on a simple theoretical framework. Consider a simple model of a motor neuron pool receiving two independent neural inputs A and B. Each independent neural input drives a different subset of motor neurons with distinct synaptic weights. As each motor neuron discharge will induce muscle fiber action potentials that will be recorded by the surface electrodes, the HDsEMG signals recorded at the muscle surface can be represented as where e(t) is a vector of the values of EMG signals sampled at time t from all available channels (less than or equal to 36 channels, after spatial filtering). eA and eB are vectors representing the EMG signals generated by the motor neurons that receive only the neural input A or B respectively and determined by the spatial distribution of the muscle fibers innervated by each motor neuron, by the (possibly overlapping) subset of motor neurons driven by each neural input, and by the synaptic strength of the neural input on each motor neuron. cA and cB are scalar values representing the level of activation of each neural input. Finally, ε(t) is the recording noise, which is assumed to be approximately constant across channels. If A and B were neural inputs only to the MNs of a single muscle, they would generate a force in the same direction, as determined by that muscle’s line of action. However, A and B may also drive different sets of additional muscles, thereby contributing to different overall limb-level force vectors through the combined action of multiple muscles. Through appropriate co-activation of antagonist and synergistic muscles, it is nevertheless possible to maintain a consistent end-point force despite differences in how A and B are distributed across muscles. In this context, a modulation of A and B that preserves the net force magnitude and direction at the endpoint can be interpreted as a muscular null space modulation. If the spatial EMG signatures associated with A and B (eA and eB) are sufficiently distinct, then the direction of the instantaneous EMG vector e(t), relative to eA and eB, could in principle provide information about this modulation. However, surface EMG signals are spatially filtered and distorted by volume conduction, surrounding tissues, and electrode configuration. As a result, the observable signatures eA and eB may be attenuated or blurred, potentially rendering them difficult to distinguish from each other and, in the limiting case, from measurement noise. To measure this modulation, we compare the vector e(t) with the vector eref which is the vector of average EMG signals recorded during trials of the Force control task. To link the control of a SuDoF to this modulation, we compute the Euclidean distance D between the vector e(t) and the vector eref: where e(t) is a vector of the values of EMG signals sampled at time t from all available channels after spatial filtering and eref is also a vector of values computed as the average of the EMG signals of each channel during trials of the Force control task and considering only moments when the force exceeded 5% of the MVF. By doing so, both the direction and the amplitude of e(t) can influence the Distance value and potentially contribute to SuDoF control. Using the Distance value for control ensures that it remains close to zero when the participant is at rest, as the magnitude of e(t) is also close to zero during this time. For an offline analysis to assess the dependence of the eref on target position, we also computed a different eref vector for each target position, as the average of the EMG signals for each channel during the Force Control task trials with the same target position, considering only samples collected when the norm of e(t) exceeded 0.2. Since this method produces multiple eref values, we interpolated eref as a function of the actual cursor position to calculate the Distance value. To convert the Distance value into a cursor angle, the 95th percentile of the Distance values from the Force control task was calculated. This value was used to define the mapping between Distance and angle such that it corresponded to a cursor rotation of 10°. Distance values were computed using a 125 ms window, during which the EMG signal was averaged to calculate e(t). Task performance To evaluate participants’ performance during the Force and SuDoF task, we calculated the Euclidean norm of the difference between the target position and the cursor position vectors (Force error), the absolute difference between the target orientation angle and cursor orientation angle (SuDoF error), and the error between both the target position and orientation and the cursor position and orientation (Total error). To compute a combined position and orientation error, the two position variables (x and y) and the orientation variable were normalized to the maximal value, and the Euclidean norm was computed from the three normalized variables. These three errors were then expressed as percentages to define three performance variables, with 0% representing the resting position or orientation, and 100% representing the target position or orientation (Force performance, SuDoF performance, Total performance). We specifically analyzed the maximum performance values reached in each trial. When the resting and target orientations were the same (i.e., 0° targets), performance was only computed when participants exerted more than 50% of the required 15% MVF (i.e., when the cursor was approximately halfway to the target position). In these trials, a SuDoF performance of 100% indicates that the cursor orientation remained at 0° while a SuDoF performance at 0% indicates that the cursor orientation was at 45°. Holding performance was evaluated as the percentage of the one-second holding period in which the cursor successfully matched the target position and orientation within the defined tolerance. To assess stability in SuDoF control, time derivative values were calculated using the ‘diff’ function in Matlab. Detection of muscular null space modulation To determine whether the Distance value could capture muscular null space modulation, we calculated the norm of e(t) over time and the angle between e(t) and eref during the Force and SuDoF control task. We used the mean value within a 250-ms window centered around the time of the highest Total performance in each trial. To avoid bias from baseline noise when the biceps were not activated, we only included angle values from a Force and SuDoF control block if the norm associated with the 0° target orientation was greater than 0.2. Identification of motor units discharge patterns To analyze motor unit activity during the Force and SuDoF control task, we decomposed the HDsEMG signals from the Holding task using a validated open-source software [31]. The Holding task was selected because its signals were more stable compared to those from the Force and SuDoF control task. For each participant, the signals from all seven contractions were decomposed and the result was manually edited [32]. Once reliable separation vectors (motor unit filters) were identified, they were applied offline over the HDsEMG signal recorded during Force and SuDoF control task using I-Spin’s function, also a validated open-source software [25]. Statistical analyses All statistical analyses were performed with RStudio using linear mixed-effect models, implemented in the R package ‘lmerTest’. Performance values were modelled as depending on the type of task (Force control task, Force and SuDoF control task with 0° trials and Force and SuDoF control task with 45° trials) as a fixed effect and participant as a random effect. When the cursor orientation approached the target orientation, it could be expected that participants would find a strategy to stabilize the cursor orientation and thus maintain it for a longer period. To verify it, we compared the time derivative of SuDoF performance between 15° and 25° and between 35° and 45°. Time derivatives of performance values were modelled as depending on the angle range (values between 15° and 25° or values between 35° and 45°) as a fixed effect and participants as a random effect. Norm and angle values were modelled depending on target orientation (0° or 45°) as a fixed effect and participants as a random effect. Distance, norm and angle computed during the Holding task were modelled depending on eref used (independent or dependent of target positions) and the type of contractions as a fixed effect and participants as a random effect. When necessary, multiple comparisons were performed using the R package ‘emmeans’, which adjusts the p-value using Tukey’s method. Normality of model residuals was checked with a Shapiro-Wilk normality test. As normality was never reached, we applied for all analyses a rank-based inverse normal transformation using the R package ‘gkeele/miqtl’. Thanks to this transformation, deviations from normality were reduced. Of note, linear models are generally robust to small violations of the normality assumption [33]. The significance level was set at p < 0.05. Values are reported as mean ± standard deviation or as the average of subject means ± between-subject standard deviation. Results Force control Before assessing performance in controlling a SuDoF through the modulation of the spatial distribution of activity in the HDsEMG electrodes of the biceps muscle, while simultaneously controlling natural DoFs, i.e., planar force exerted at the hand, we first considered force control performance. In both the Force control task and in the Force and SuDoF control task for trials where the target orientation matched the resting position, only force control was required. Examples of the control of force during these control conditions are shown in Fig. 3. In the Force control task (Fig. 3A), participants successfully reached the target in 87 ± 14% of the trials but managed to remain on the target for the required time (1 s) in only 34 ± 21% of the trials. This indicates that, while participants had little difficulty reaching the target, holding the position proved more challenging. Similar results were observed during the Force and SuDoF control task for trials where the target orientation matched the resting position (Fig. 3B, left), considering only force performance, where the success rates for reaching and holding the target were 94 ± 7% (Fig. 4A, left) and 39 ± 14% (Fig. 4D, left), respectively. Significant differences were found between the Force control task and the Force and SuDoF control task for trials with the same target orientation as the resting position, in terms of Force performance and Holding performance (p < 0.001 for both). The performance improvement observed during the Force and SuDoF task could be attributed to the lower force amplitude required to reach the targets (15% of MVF) compared to the Force control task, which required 10%, 20% or 30% of MVF. The fact that the increase was also observed for Holding performance, despite the additional requirement to maintain the cursor orientation at 0°, indicates that setting the 95th percentile of the distribution of the Distance values observed in the Force control task to correspond to 10° of cursor rotation effectively prevented participants from being disrupted while focusing solely on Force control. Force control task and Force and SuDoF control task examples. A Cursor trajectories during the Force control task are shown for participant 1 and participant 16 in different trials with different target positions and target force magnitudes. Trajectories are color-coded according to the target position, indicated by a red cross. B Cursor trajectories and SuDoF performance over time during the Force and SuDoF control task. All trials from the fifth block of Force and SuDoF control task are shown for participant 1 (top) and participant 16 (bottom), who was unable to rotate its cursor during the task. For each participant, trials with a target orientation at 0° are displayed on the left. Cursor trajectories are shown in the top part of the figure, with the target position marked by a red cross. Cursor orientations are displayed in the bottom part, showing SuDoF performance over time, with success thresholds indicated by red dashed lines. The 95th percentile of the distribution of Distance values from the Force control task is marked by a red dotted line. As the 95th percentile corresponded to a cursor orientation of 10° and the target orientation was 0° or 45°, it represents a SuDoF performance of 22% for a target at 45° and of 78% for a target at 0° Force and SuDoF control performance. Each panel shows one of the four performance metrics during the Force and SuDoF control task with target orientations at 0° and 45°. The maximum of each performance metric in each trial is averaged over all trials of each participant. Each participant is represented by a circular marker of a different color. Boxplots are shown for each target orientation, with the boxes representing the first and third quartiles, whiskers showing the maximum and minimum non-outlier values, and the red line indicating the median. Success thresholds are indicated with red dashed lines, and the 95th percentile of Distance values from the Force control task is indicated by a red dotted line. As the 95th percentile corresponded to a cursor orientation of 10° and the target orientation was 0° or 45°, it represents a SuDoF performance of 22% for a target at 45° and of 78% for target at 0°. The left side of each graph shows trials where the target was at 0° while those with a 45° target are shown on the right. A Normalized error between target position and cursor position (Force performance). B Normalized error between target orientation and cursor orientation (SuDoF performance). C Normalized error combining both the position and orientation errors (Total performance). D Percentage of the one-second holding period in which the cursor was successfully maintained within the target (Holding performance) We then considered Force performance in the Force and SuDoF control task for trials with target orientation at 45° (Fig. 3B, right). In these trials, which also required SuDoF control, the success rate for reaching the target was 89 ± 10% (Fig. 4A, right). Significant differences were found for the Force performance between the Force control task and the Force and SuDoF control task for trials with target orientation at 45° (p = 0.005). The improved Force performance observed could likely be attributed to the lower force amplitude required. Of note, Force performance was not significantly different between Force and SuDoF control trials with target orientation at 0° and at 45° (p = 0.19). Overall, attempting to control the SuDoF did not impair force control, and participants remained able to control their natural DoFs. Force and Holding performance data for individual trials of each participant are presented in Additional file 1 (Figure S1). Simultaneous force and muscular null space control The SuDoF, i.e., cursor rotation, was controlled by computing a Distance value in the space of HDsEMG signals, which measured modulation within the muscular null space. To minimize unintentional cursor rotation, the mapping of the Distance value to cursor rotation was set so that a 10° rotation of the cursor corresponded to the 95th percentile of the distribution of Distance values from the Force control task, minimizing the occurrence of unwanted rotations. In the Force and SuDoF control task all the participants successfully maintained the Distance value below the 95th percentile of the distance distribution in Force control in all the trials with 0° target orientation, i.e. the same cursor orientation maintained at the resting position. However, reaching a 45° cursor orientation was more challenging (Fig. 3B, right). While in 93 ± 9% of the trials with 0° target orientation participants reached a SuDoF performance of at least 90% (Fig. 4B, left), this occurred in only 39 ± 32% of the trials with target orientation at 45° (Fig. 4B, right). In most of the trials, participants could not reach the target orientation at 45°. And if they were able to reach the target orientation, they never successfully maintained both position and orientation for the required time (Fig. 4D, right). However, they were able to increase the Distance value above the 95th percentile of the distance distribution in Force control in 84 ± 31% of the trials (Fig. 4B right). More specifically, thirteen participants could systematically successfully increase the Distance value above the 95th percentile (in 98 ± 2% of the trials) while three participants could do so in only 20 ± 4% of the trials (Fig. 4B right). On average, these thirteen participants showed a SuDoF performance of 79 ± 23%. SuDoF performance and Total performance data for individual trials of each participant are presented in Additional file 1 (Figure S1). Figure 3B shows that as the Distance value increases, its variability also increases, with large fluctuations appearing, highlighting the difficulty in precisely controlling it. This is further reflected in the Holding performance (Fig. 4D), as in no trials participants could successfully hold the cursor at the target orientation of 45°. To investigate this further, we compared the time derivative of SuDoF performance between 15° and 25° and between 35° and 45°. The derivative values were significantly smaller around 20° than around 40° (p < 0.001). Given that the required target orientation was almost reached at 40°, it would have been beneficial for participants to find a strategy to stabilize the cursor orientation and thus maintain it for a longer period. However, the large variability observed at 40° indicates the difficulty participants had in maintaining this orientation and suggests that a target with a smaller angle would have improved Holding performance. To better understand the strategies employed by participants, we considered the force and torque generated during the Force and SuDoF control task (Fig. 5). As expected, to succeed in the task, participants did not change horizontal force production (Fxy, Fig. 5A) between trials with a 0° target and trials with a 45° target (0.15 ± 0.00% and 0.15 ± 0.01% of MVF respectively, p = 0.32). However, they significatively increased the vertical force production (Fz, Fig. 5B) between trials with a 0° target and trials with a 45° target (0.02 ± 0.01% and 0.03 ± 0.01% of MVF respectively, p < 0.001) and supination or pronation torque (TorqueY, Fig. 5C, 0.10 ± 0.05% and 0.22 ± 0.15% of MVT respectively, p < 0.001). Most participants successfully controlled cursor orientation without modifying their horizontal force production. Force and torque data for individual trials of each participant are presented in Additional file 1 (Figure S2). Horizontal and vertical forces and torque during Force and SuDoF control. A Horizontal force (affecting cursor position), B vertical force (not affecting cursor position), and C torque around the Y-axis (associated to wrist supination or pronation) during the Force and SuDoF control task. For each participant and trial, mean values within a 250-ms window centered around the point of best Total performance are computed. Averages across trials of each participant are represented by small circles. Trials with a 0° target orientation are shown on the left, while those with a 45° target orientation are shown on the right. Values from the same participant are connected by a line. The mean values of trials with 0° and 45° target orientations across all participants are represented by large black circles. The mean values of trials during the Force control task are indicated by a red horizontal line for comparison. MCT: maximum voluntary torque. Detection of multiple neural inputs to the biceps brachii According to our theoretical framework, if a muscle receives only a single neural input, the vector containing instantaneous HDsEMG signal values [e(t)] should change its direction only because of noise. However, if multiple neural inputs drive different subsets of motor neurons within the pool of a single muscle and if the activations of those subsets generate different patterns of HDsEMG activity, then the direction of e(t) is expected to vary when the two inputs are modulated across conditions. Such changes should already be detectable during the Force control task, as participants produced force in different directions. Thus, we compared the mean HDsEMG patterns during each trial of the Force task by computing the angle between e(t) during pairs of trials (Fig. 6). We considered only e(t) with a norm exceeding 0.2 to discard noisy patterns when the biceps muscle was not activated. We found that angles between trials with different target positions but with the same target force magnitude were significantly higher than angles between trials with the same target position but differing in force magnitude (18 ± 6° and 10 ± 2°, respectively, p < 0.001) or angles between trials that are a repetition of the same condition (8 ± 3°, p < 0.001). Angles between trials differing in force magnitude were also significantly higher than angles between trials with the same condition (p < 0.001). Angle data for individual trials of each participant are presented in Additional file 1 (Figure S3). Together, these results are compatible with the presence of multiple neural inputs to the biceps brachii, reflected in systematic variations of HDsEMG patterns across task conditions. Angles between mean HDsEMG patterns of pairs of trials during the Force control task. Each boxplot represents the angles between the mean HDsEMG patterns of pairs of trials, averaged across all pairs of each subject (P1 to P16, identified by the marker color as in Fig. 8), in each one of three conditions: (i) trials with the same target position but differing in force amplitudes (orange box plot), (ii) trials with different target positions but with the same force amplitude (yellow) and (iii) repeated trials with the same target position and force amplitude (purple). The boxes represent the first and third quartiles, the red line marks the median, and the whiskers extend to the maximum and minimum non-outlier values If multiple neural inputs drive the motor neurons of the biceps, participants could use them to control the SuDoF. The Distance value controlling the SuDoF depends on both the norm of e(t) and the angle between e(t) and eref, the reference vector. Participants could then control the SuDoF by varying the neural inputs without increasing the total drive to the motor neurons, and thus by mainly modifying the angle between e(t), or by also increasing the norm of e(t). Figure 7 reports the norm of e(t) and the angle between e(t) and eref during the Force and SuDoF control task. We found a significant difference in the norm of e(t) between trials with a 0° target and trials with a 45° target (0.56 ± 0.24 and 3.5 ± 2.2 respectively, p < 0.001). This indicates that participants generated the same force output with a sevenfold increase in muscle activation. We also found a small but significant difference in the angle between e(t) and eref when comparing trials with a 0° target and trials with a 45° target (16 ± 3° and 18 ± 3° respectively, p < 0.001). Norm and angle data for individual trials of each participant are presented in Additional file 1 (Figure S4). Thus, participants controlled the SuDoF mainly by increasing the norm of e(t) but also by slightly increasing the angle between e(t) and eref. Norm and angle of e(t) during Force and SuDoF control. A Norm of e(t) and B angle between e(t) and eref averaged over time during the Force and SuDoF control task and trials performed by each participant. For each trial, mean values are computed within a 250-ms window centered around the time of best Total performance. For each participant, small circles indicate the average over trials. Trials with a 0° target orientation are shown on the left, while those with a 45° target orientation are shown on the right. Average values for trials of the same participant are connected by a line. To avoid bias from baseline signals when the biceps were not activated, only angle values from Force and SuDoF control blocks where the norm associated with the 0° target orientation exceeded 0.2 were included. Mean values across all participants for trials with 0° and 45° target orientations are represented by large black circles To confirm that multiple neural inputs contribute to the increase in the Distance value, we analyzed the HDsEMG signals from the Holding task by decomposing them to examine motor unit discharge patterns. In total, only 56 motor units were decomposed with an average of 3.5 ± 4.5 units per participant (ranging from 0 to 13). After identifying reliable motor unit filters, they were applied offline over the HDsEMG signal recorded during the Force and SuDoF control task. However, no motor units could be reliably tracked during the Force and SuDoF control task. This failure may be attributed to the substantial increase in muscle activation during SuDoF control. These findings highlight the advantage of the Distance value: it is consistently applicable across participants and is less reliant on signal quality or reference task parameters. To verify that multiple neural inputs can be detected by our new method, we analyzed the norm of e(t), the angle between e(t) and eref, and the Distance value during the Holding task. In this task, participants performed and sustained contractions targeting different functions of the biceps brachii, increasing the likelihood of engaging distinct neural inputs for each contraction. Using the HDsEMG signals recorded during this task, we calculated the Distance value, the norm of e(t), and the angle between e(t) and eref (Fig. 8). Since the analysis of the angle between e(t) during pairs of trials of the Force task (Fig. 6) indicated that HDsEMG patterns depend on the target position, we computed these values using two approaches: an eref independent of the target position (as used to compute the Distance during the experiment) and an eref dependent on the target position (derived as the mean HDsEMG signal from trials with the same target position). Interestingly, the norm and angle were not significantly different between values computed with eref dependent or independent of the target position (p = 0.97 and p = 0.42, respectively). However, the Distance value showed a significant difference (p < 0.001). Specifically, the SuDoF performance during supination averaged 21 ± 21% and 35 ± 42% for eref independent and dependent on the target position, respectively. During co-contraction, the SuDoF performance averaged 18 ± 11% and 38 ± 34%. While the norm of e(t) remained similar when eref was modified, the SuDoF performance increased significantly for contractions like supination and co-contraction. This is because even a small modification of the angle between e(t) and eref, when combined with an increase in norm, led to a significant increase in the Distance value. Distance value, norm and angle of e(t) during Holding task. A The mean Distance value during the final 15 s of each sustained target in the Holding task, computed for each participant. To allow comparison, the distance is normalized and displayed as SuDoF performance value with 100% corresponding to the target orientation of 45°. The red dotted line represents the 95th percentile of the distribution of Distance values during the Force control task. Orientation success thresholds are marked with red dashed lines. B The mean norm of e(t), computed during the final 15 s of each sustained target in the Holding task for each participant. C The mean angle between e(t) and eref was computed during the final 15 s of each sustained target in the Holding task for each participant. Boxplots are shown for each sustained target, with the boxes representing the first and third quartiles, whiskers showing the maximum and minimum non-outlier values, and the red line indicating the median. Individual data points are shown, each represented by a participant-specific colored marker. Orange boxplots (left) display values calculated using an eref independent of the target position (as used during the experiment). Yellow boxplots (right) show values calculated with an eref dependent on the target position. For the latter, eref was derived as the mean HDsEMG signal from trials with the same target position and a norm greater than 0.2, with eref then interpolated as a function of the actual position to compute the Distance value Discussion We introduced a method to quantify modulations within the HDsEMG space that do not affect the force output and to map these modulations onto the control of a SuDoF during a motor task requiring the control of natural DoFs, i.e. the generation of multidirectional isometric force to control the position of a cursor. The method compares ongoing EMG patterns with a reference pattern obtained during natural arm control and quantifies their deviation using a Distance metric in the space of HDsEMG signals. Participants were able to use this signal to control the rotation of the cursor while maintaining horizontal force production. Although the Distance metric depends on both the direction and magnitude of HDsEMG vector, participants mainly relied on increases in magnitude to control the SuDoF, with comparatively smaller changes in direction. Nevertheless, directional differences were consistently present and increased when task conditions differed, suggesting that the HDsEMG patterns were not modulated solely through scaling of a single neural input. These directional variations are compatible with changes in the relative contributions of distinct neural inputs to different subpopulations of the motor neuron pool, although our results do not allow a direct identification of distinct neural inputs at the motor neuron level. Rather, the findings indicate that task-relevant information can be extracted from structured modulations in HDsEMG patterns without requiring motor unit decomposition. Theoretical framework validation Our framework assumes that partially independent neural drives may project to overlapping subpopulations within a motor neuron pool, potentially giving rise to distinguishable spatial EMG patterns. In practice, however, the relationship between neural input and HDsEMG distribution is shaped by motor unit recruitment [34], nonlinear summation, volume conduction, and noise. As recruitment increases, spatial EMG patterns may change due to an increase in noise and to the progressive activation of motor units with different spatial signatures. To validate our framework against alternative interpretations, we investigated whether the Distance metric was merely a byproduct of noise or simple spatial recruitment. Offline analysis revealed that the metric is robust to noise, as increasing the signal filtering window from instantaneous to 500 ms did not suppress SuDoF performance (Figure S5 in Additional file 1). Furthermore, while spatial recruitment—the progressive activation of motor units as force increases—does influence the HDsEMG pattern, our results show that the angular shifts associated with force magnitude changes (Fig. 6, “Amplitude”) are significantly smaller than those observed during changes in force direction (Fig. 6, “Direction”) or SuDoF control. This suggests that while recruitment contributes to the Distance value, it cannot fully account for the structured modulations we observed. This observation partially aligns with findings by Borzelli et al. [35], who reported that contraction level influenced the spatial distribution of motor unit recruitment in the biceps brachii. However, given that HDsEMG provides an indirect measure of motor neuron activity and that our method does not rely on motor unit decomposition [26], the present results should be interpreted cautiously. Rather than directly identifying distinct neural inputs, our findings indicate that structured variations in HDsEMG patterns, consistent with the possibility of partially differentiated neural drives within one muscle, can be detected and exploited for control without explicit motor unit analysis. Dependence on the reference pattern and fatigue The Distance metric was strongly dependent on the definition of the reference pattern (eref). When eref was not adapted to target position, null-space modulation during the Holding task was difficult to distinguish from baseline variability (Fig. 8). Adapting eref to the task context improved sensitivity, particularly for supination and co-contraction conditions. This highlights an important limitation: null-space detection using this method relies critically on how reference patterns are defined and updated. In practical applications, adaptive or task-specific calibration procedures may therefore be necessary to ensure robustness. Fatigue further influenced HDsEMG patterns, as reflected by the higher activation levels and altered angular relationships observed during combined Force and SuDoF control. Because the method detects deviations relative to eref, any physiological changes that alter EMG distributions, such as fatigue [36], can affect the Distance metric independently of intentional null space modulation. Alternative explanations and task-level factors Potential alternative strategies for controlling the SuDoF included excessive supination, vertical force production, or co-contraction. Although small increases in vertical force and torque were observed, their magnitude was insufficient to explain the substantial increase in EMG amplitude during SuDoF control. The dominant feature was a marked rise in overall muscle activation, consistent with co-contraction. From a task-level perspective, variability in dimensions not directly constrained by the goal (e.g., vertical force) may naturally increase [37], but these factors alone do not account for the observed Distance modulation. Crosstalk from surrounding muscles cannot be entirely excluded, given the use of surface EMG. However, spatial filtering and the relatively small angular deviations suggest that widespread contamination was unlikely to be the primary driver of the control signal. Importantly, even if multiple muscles contributed to the modulation, maintaining the required endpoint force would still necessitate coordination consistent with a task-level null space control strategy. Control performance and practical limitations Participants successfully controlled the SuDoF while maintaining horizontal force, primarily by increasing EMG amplitude relative to eref. Because eref was not target-specific, a baseline angular difference already existed between eref and the task-related EMG patterns, making amplitude modulation an efficient control strategy. However, achieving the 45° cursor orientation required Distance values substantially exceeding those observed during force control alone. This likely contributed to the increased variability and difficulty in maintaining that target, particularly under fatigue. Adjusting the control gain or redefining the reference space may allow more stable and less effortful operation in future implementations. Conclusions Overall, the results demonstrate that structured modulations within HDsEMG space can be used to generate an additional control signal concurrent with a motor task involving natural DoFs. While the data are compatible with the presence of partially differentiated neural drives within a single muscle, the current evidence supports a more conservative interpretation: the method detects reproducible, task-related modulations in spatial EMG patterns that can be exploited for control, without requiring explicit motor unit decomposition. This approach may provide a practical framework for augmenting motor output using surface EMG, provided that reference definition, fatigue effects, and control gain are carefully managed. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1: Description: This file contains detailed results for each individual participant (P1–P16) to supplement the population-level data presented in the main manuscript and the result of an off-line analysis performed to assess the contribution of noise to SuDoF performance. It includes: Figure S1: Force, SuDoF, Total, and Holding performance for all trials of each participant. Figure S2: Horizontal/vertical forces and torque during Force and SuDoF control for all trials of each participant. Figure S3: Angles between mean HDsEMG patterns of pairs of trials during the Force control task for each participant. Figure S4: Norm and angle of e(t) during Force and SuDoF control for all trials of each participant. Figure S5: Effect of filtering of HDsEMG signals on SuDoF performance. Acknowledgements Not applicable. Abbreviations Degrees-of-freedom Electromyography High-density surface EMG Maximal voluntary contractions Maximal voluntary force Maximal voluntary torque Supplementary degrees-of-freedom Author contributions JR and AdA conceptualized and designed the study. JR acquired and analyzed the data. JR and AdA interpreted the data and drafted the work. All authors read, substantively revised, and approved the final manuscript. Funding This study was funded by the European Union’s Horizon Europe programme project “HARIA: Human-Robot Sensorimotor Augmentation - Wearable Sensorimotor Interfaces And Supernumerary Robotic Limbs For Humans With Upper-Limb Disabilities”, under grant agreement No. 101070292. JR and EH were supported by the HARIA project. Data availability The processed data analyzed for the study are available in the HARIA Zenodo repository (doi: https://zenodo.org/records/20387300). Declarations Ethics approval and consent to participate Participants gave written informed consent. All procedures have been conducted in accordance with the principles embodied in the Declaration of Helsinki, comply with national regulations, and have been approved by the Lazio territorial ethics committee Area 5 (N.82/SL/23). Consent for publication Not applicable. Competing interests JR and AdA have filed a patent application related to the method described in this manuscript. All other authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1: Description: This file contains detailed results for each individual participant (P1–P16) to supplement the population-level data presented in the main manuscript and the result of an off-line analysis performed to assess the contribution of noise to SuDoF performance. It includes: Figure S1: Force, SuDoF, Total, and Holding performance for all trials of each participant. Figure S2: Horizontal/vertical forces and torque during Force and SuDoF control for all trials of each participant. Figure S3: Angles between mean HDsEMG patterns of pairs of trials during the Force control task for each participant. Figure S4: Norm and angle of e(t) during Force and SuDoF control for all trials of each participant. Figure S5: Effect of filtering of HDsEMG signals on SuDoF performance. Data Availability Statement The processed data analyzed for the study are available in the HARIA Zenodo repository (doi: https://zenodo.org/records/20387300).