Fifteen young non-impaired adults (Table 1) consented to participate in this institutionally-approved study. Participants were free of neurological, cardiovascular, or musculoskeletal disorders, and able to walk independently. Individuals with any lower extremities’ injury within 6 months before the participation were excluded.

We used a dual-belt, force-instrumented treadmill equipped with side hand railings (Motek, Amsterdam, BG). A ten-camera motion capture system (Vicon, Oxford EN) tracked passive reflective markers, that we processed offline (Vicon Nexus) to create a 12-segment rigid body biomechanical model. Real-time visual feedback was generated in Vicon Nexus using the built-in Monitor tool, which displayed the GRFs on a step-by-step basis and streamed the output to a graphical display. The feedback was delivered on a white screen positioned approximately 1.5 meters in front of the participant at eye level. The interface consisted of a vertical line that updated each step to represent the GRFs of the non-dominant limb, along with a horizontal target line overlaid with a shaded region indicating the participant-specific AGRF target (5% or 10% body weight). Participants were instructed to modulate their force output so that the bar reached the target line at each step (Figure 1). To create the backward directed resistive forces, we utilized a custom-built passive device that applied the resistive force through a belt won at the level of the center of mass (COM) (Figure 2). The custom device used rubber tubing fed through a winch to change force levels, which was measured with a load cell (RAS1-250S-S*C02, Loadstar Sensors Inc., Fremont, CA, USA) in series with the tubing. For safety, participants wore a fall harness that provided no support during treadmill walking but protected them if they lost their balance during the protocol.

Participant demographics, including weight, and dominant leg, were documented at the beginning of the session. Limb dominance was determined by participant self-report. Participants were asked which leg they naturally prefer for strength or balance tasks, like kicking a ball. Comfortable Walking Speed (CWS) was first measured overground on a Zeno gait mat (5.5 m × 0.91 m) and analyzed using ProtoKinetics Movement Analysis Software (ProtoKinetics, Havertown, PA). Participants completed two straight line walking passes across the walkway at their CWS. The instruction provided was: “Walk at a speed that feels most comfortable to you.” To ensure steady-state gait within the capture area on the mat, participants began walking 2 meters before the mat and continued 2 meters beyond it. After completing the first pass, they stopped, reset, and repeated the procedure for the second pass using the same instructions. Subsequently, each participant’s CWS on the treadmill was determined through a standardized protocol. Initially, the treadmill speed was set approximately 0.3 m/s slower than the participant’s overground speed and gradually increased in increments of 0.1 m/s until the participant reported, “This feels like my normal walking speed.” This speed was recorded, and the procedure was repeated starting at a speed faster than the previously identified speed, gradually decreasing by increments of 0.1 m/s until the participant again indicated their normal walking speed. The average of these two treadmill speeds was calculated and defined as the participant’s CWS for the treadmill condition (Dal et al., 2010; Martin et al., 1992).

Next, participants walked on the treadmill at their CWS for 1 min to obtain baseline measurements. Target Anterior Ground Reaction Forces (AGRFs) for visual feedback were then determined. Participants walked on the treadmill at half their CWS for 30 s, and the data was exported from Vicon Nexus software. This data was then analyzed using a custom script in MATLAB (MathWorks Inc., Natick, MA) to calculate the mean peak AGRFs for the non-dominant leg. The 5% visual feedback target AGRF was determined by adding 5% of the participant’s body weight to the mean peak AGRF of the non-dominant leg. The same procedure was followed to establish the 10% visual feedback target AGRF.

After establishing the target AGRFs, participants completed an adaptation trial on the split-belt treadmill. The non-dominant leg was placed on the slow belt (set at half of the participant’s CWS), and the dominant leg was placed on the fast belt (set at the participant’s CWS). The adaptation trial lasted between 3 and 6 min. Initially, all participants completed a standardized 3-min adaptation period. After that, participants were asked whether they felt stable and comfortable, and the research team visually assessed gait for signs of instability or a lack of gait smoothness. For a few participants who reported continued discomfort or demonstrated observable instability, the adaptation time was extended in one-minute increments. Reassessment occurred after each minute, with the adaptation period ending once a stable walking pattern was observed. No participant required more than 6 min. This individualized approach ensured adequate acclimation to the split-belt condition before experimental testing.

Following this adaptation trial, participants completed an experimental protocol consisting of 6 trials, grouped into two feedback conditions (Intrinsic via backward resistive force and extrinsic via visual feedback). The order of feedback conditions was randomized to minimize adaptation and maintain engagement while target levels were delivered in a fixed ascending order within each feedback type. In addition, rest periods were provided between trials.

Before data collection, participants completed a familiarization with the visual interface while walking on the treadmill. During this orientation, they were shown how the GRF’s line height changed with their push-off force and limb position, and they practiced reaching the target line to ensure they understood the task. Participants were instructed to position their non-dominant leg behind their hip (increasing the trailing limb angle) and to push harder on the treadmill using their plantar flexors without holding the handrail. A similar familiarization was provided before the backward resistive force (BRF) condition, during which participants experienced the applied resistive load and were instructed on how to maintain their position at the center of the treadmill without using the handrails but use their legs to overcome the forces.

Intrinsic feedback: participants walked on the split-belt treadmill with the dominant leg on the fast belt at CWS and the non-dominant leg on the slow belt at half CWS, with backward resistive forces applied to the COM at 0, 5, and 10% of body weight (BW), respectively (Figure 2).

Extrinsic feedback: participants walked on the split-belt treadmill with the dominant leg on the fast belt at CWS and the non-dominant leg on the slow belt at half CWS while receiving real-time AGRF visual feedback at targets of 0% (no target), 5, and 10% of BW, respectively (Figure 1).

Kinematic data was collected at 100 Hz using the capture motion system, while kinetic (ground reaction forces) were collected at 1000 Hz via the treadmill’s force plates. Both ground reaction forces and segment kinematic data were filtered in Vicon Nexus software (version 2.15, Vicon Motion Systems, Oxford, UK) using a 6 Hz and 12 Hz cutoff frequency for kinematics and kinetics, respectively. The filtered data were then exported for post-processing to Visual 3D (C-motion, Germantown, MD) and MATLAB (MathWorks Inc., Natick, MA).

For non-dominant limb (on slow belt), the peak AGRF was identified for every step as the maximum value of the anterior-directed force component (Fy) during the stance phase, defined from heel strike to toe-off with a threshold of 10 N. The mean peak AGRF was then calculated for the limb by averaging the peak propulsion values across all steps within each trial. After computing the peak anterior ground reaction force (AGRF) for each trial, we calculated the difference from the target AGRF by subtracting the mean peak AGRF of the leg on the slow belt from the predefined target value. This difference represented the participant’s error in achieving the target propulsion during each feedback condition.

Trailing limb angle was calculated as the angle between the vertical axis and the vector connecting the greater trochanter and lateral malleolus (ankle marker) in the sagittal plane at toe-off (Miyazaki et al., 2019). Toe-off was determined based on when the vertical ground reaction force dropped below 10 N, indicating the end of the stance phase and the start of the swing phase.

An inverse dynamic routine was used to quantify sagittal plane joint moments and positive powers at the hip and ankle, utilizing commercial software (Visual 3D, C-Motion Inc., Germantown, MD). For each step, the moments and positive power were calculated during the stance phase (from heel strike to toe-off) and normalized to the individual’s body mass.

Both intrinsic and extrinsic feedback conditions yielded similar propulsive force outputs. The average error from the target was summarized as mean ± SD at each of the two target levels. There was no significant main effect of feedback on average propulsion error (F (1, 14) = 0.20, p = 0.66) with a small effect size (ηp² = 0.014), indicating that participants achieved similar accuracy in reaching the target propulsion level under both visual and backward resistive force (BRF) conditions. At the 5% target level, the mean error was 0.14 ± 0.18 N/kg for visual feedback and 0.18 ± 0.17 N/kg for BRF. At the 10% target level, the mean was 0.00 ± 0.24 N/kg for visual feedback and 0.03 ± 0.22 N/kg for BRF. These findings suggest that participants were able to match the target propulsion level similarly under both feedback conditions (Figure 3).

Trailing limb angle was moderated by type of feedback and amount of force. As shown in Figure 4, we found significant feedback × target interaction (F (2, 28) = 15.76, p = 0.0002), indicating that the effect of target level on TLA depended on feedback type. Post hoc comparisons with Bonferroni correction revealed that, at each target level, visual feedback produced significantly greater TLA than BRF. At the 5% target, TLA was 22.42 ± 2.61° for visual and 20.51 ± 3.14° for BRF (t (80, 35) = −3.39, p = 0.0011) with a medium effect size (η² = 0.12). At the 10% target, TLA was 25.30 ± 3.23° for visual and 21.82 ± 2.97° for BRF (t (80, 35) = −6.18, p < 0.0001) with a large effect size (η² = 0.21). In addition, a significant main effect of target level was observed (F (2, 28) = 72.88, p < 0.0001), with TLA increasing incrementally across the 0, 5, and 10% targets. A main effect of feedback was also present (F (1, 14) = 22.04, p = 0.0003), with visual feedback producing greater TLA overall compared to BRF. Descriptive results (mean ± standard deviation) are presented in Figure 5.

Ankle moment was moderated by type of feedback and amount of force. As shown in Figure 6, significant feedback × target interaction was observed (F (2, 28) = 5.65, p = 0.0087), indicating that the effect of target level on ankle moment differed depending on feedback type. In addition, a significant main effect of target level was observed (F (2, 28) = 27.19, p < 0.0001), with ankle moment increasing incrementally as propulsion targets increased. A main effect of feedback was also present (F (1, 14) = 16.64, p = 0.0011), with visual feedback producing greater ankle moments overall compared to BRF. Post hoc comparisons with Bonferroni correction revealed that visual feedback elicited significantly greater ankle moment than BRF at both the 5 and 10% target levels. At the 5% target, the mean ankle moment was 1.29 ± 0.21 Nm/kg for visual feedback and 1.16 ± 0.29 Nm/kg for BRF (t (80, 35) = −3.63, p = 0.0005) with a medium to large effect size (η² = 0.14). At the 10% target, ankle moment was 1.36 ± 0.15 Nm/kg for visual feedback and 1.20 ± 0.25 Nm/kg for BRF (t (80, 35) = −4.69, p < 0.0001) with a large effect size (η² = 0.21). These results suggest that participants used more ankle joint moments under visual feedback. Descriptive results (mean ± standard deviation) are presented in Figure 5.

Hip moment was moderated by type of feedback and amount of force. As shown in Figure 7, a significant feedback × target interaction was observed (F (2, 28) = 14.76, p = 0.0004), indicating that the way hip moment changed with target level depended on feedback type. In addition, a significant main effect of target level was observed (F (2, 28) = 17.22, p < 0.0001), with hip moment increasing as propulsion targets increased. A main effect of feedback was also present (F (1, 14) = 11.88, p = 0.0039), with BRF producing greater hip moments overall compared to visual feedback. Post hoc comparisons with Bonferroni correction revealed that BRF elicited significantly greater hip moment than visual feedback at both the 5 and 10% target levels. At the 5% target, the hip moment was 0.64 ± 0.13 Nm/kg for BRF and 0.56 ± 0.15 Nm/kg for visual feedback (t (80, 35) = 3.23, p = 0.0018) with a medium effect size (η² = 0.11). At the 10% target, hip moment increased to 0.70 ± 0.12 Nm/kg under BRF and 0.55 ± 0.14 Nm/kg under visual feedback (t (80, 35) = 6.47, p < 0.0001) with a large effect size (η² = 0.34). Descriptive results (mean ± standard deviation) are presented in Figure 5.

Ankle power was moderated by type of feedback and amount of force. As shown in Figure 8, significant feedback × target interaction was present (F (2, 28) = 7.94, p = 0.0019), indicating that the effect of feedback on ankle power varied depending on the propulsion target level. In addition, a significant main effect of target level was observed (F (2, 28) = 8.34, p = 0.0050), with ankle power increasing as propulsion targets increased. There was no significant main effect of feedback (F (1, 14) = 5.19, p = 0.0389). Post hoc comparisons with Bonferroni correction revealed that, at 5%, there was no significant between feedback types (t (80, 35) = −0.31, p = 0.0754). The average ankle power was 1.29 ± 0.59 W/kg under visual feedback and 1.26 ± 0.42 under BRF. However, at 10%, the average ankle power was 1.20 ± 0.39 W/kg under visual feedback and 1.55 ± 0.53 under BRF. Ankle power was significantly greater under BRF compared to visual feedback (t (80, 35) = 4.73, p < 0.0001) with a large effect size (η² = 0.21). These results suggest that BRF enhances ankle power generation specifically when the propulsion demand is high. Descriptive results (mean ± standard deviation) are presented in Figure 5.

Hip power was moderated by type of feedback and amount of force. As shown in Figure 9, significant feedback × target interaction was present (F (2, 28) = 27.11, p < 0.0001), indicating that the effect of feedback on hip power varied depending on the propulsion target level. In addition, a significant main effect of target level was observed (F (2, 28) = 61.51, p < 0.0001), with hip power increasing incrementally as propulsion targets increased. A main effect of feedback was also present (F (1, 14) = 45.43, p < 0.0001), with BRF producing greater hip power overall compared to visual feedback. Post hoc comparisons with Bonferroni correction revealed that, at the 5% target, hip power was 0.50 ± 0.13 W/kg under BRF and 0.28 ± 0.11 W/kg under visual feedback. This difference was statistically significant (t (80, 35) = 6.45, p < 0.0001) with a large effect size (η² = 0.34). At the 10% target, hip power further increased under proprioceptive feedback to 0.64 ± 0.13 W/kg, while it remained lower under visual feedback at 0.34 ± 0.17 W/kg. This difference was also significant (t (80, 35) = 9.02, p < 0.0001) with a large effect size (η² = 0.50). Descriptive results (mean ± standard deviation) are presented in Figure 5.

The current study has several limitations that should be discussed. First, we acknowledge that our estimation of CWS may not fully reflect each participant’s true overground preferred walking speed. However, because this study used a repeated-measures design in which all feedback conditions (intrinsic and extrinsic) were tested relative to each participant’s own treadmill-derived CWS, any potential inaccuracy would have influenced all conditions equally. Thus, we believe our primary outcomes remain comparable across conditions. Second, during familiarization with visual feedback, participants were instructed to place their non-dominant leg behind their hip to help them reach the AGRF target, but not in BRF, as reflected in TLA results. Regardless, these differences likely contributed to the distinct biomechanical strategies observed but did not affect the primary outcome. Participants achieved similar propulsion accuracy across conditions. Third, no neural-level measurements, such as EMG, cortical and subcortical imaging, etc., were collected in this study. Although our biomechanical variables capture the primary mechanical contributors to propulsion, they do not allow direct inference about the underlying neural control processes. Nevertheless, since our interpretations focus on mechanical strategy selection rather than neural control mechanisms, the lack of neural-level measurements does not affect our primary conclusions. In addition, our analysis was limited to propulsion strategies quantified through TLA and ankle and hip joint kinetics. Although these variables represent the primary mechanical pathways for generating forward propulsion, they do not capture knee joint contributions, segmental mechanical work, or muscle activation patterns. However, the group-level results clearly showed different biomechanical strategies across feedback conditions while participants achieved similar levels of propulsion, even without measuring all possible propulsion contributors. Participants walked under the split-belt paradigm for only 3 to 6 min in the adaptation trials, which may not have allowed complete adaptation of the gait pattern to reach a steady state. Previous studies have shown that locomotor adaptation, especially in step length symmetry, can continue for a long time in a novel split-belt paradigm (Sánchez et al., 2019). However, differences become relatively small after initial strides of adaptation (Reisman et al., 2007), and we have previously shown that kinetic measures can adapt quickly during short split-belt exposures (Hurt et al., 2022). Therefore, while prolonged walking may reveal additional adaptations, our short adaptation trials were sufficient to capture immediate joint-level strategies under feedback. The duration of gait trials with feedback was short (1 min) and differences between the types of feedback may potentially arise during longer-duration walking trials. However, our question focused on the immediate effect of feedback rather than long-term adaptation. Split-belt treadmill and feedback studies show that changes in different gait outcomes and propulsion measures can occur within minutes or even on the initial exposure (Vasudevan and Bastian, 2010; Reisman et al., 2010; Liu et al., 2020). Lastly, we limited assessing outcomes to the slow belt only of the split-belt walking trial. However, several studies have already shown how these types of feedback influence propulsion and joint-level strategies during tied-belt walking (Kuhman and Hurt, 2019a; Browne and Franz, 2018). Therefore, our conclusions were restricted to the split-belt walking paradigm by design.

This study provides insight into how individuals with a non-impaired nervous system can generate similar peak AGRF to achieve different target levels during split-belt treadmill walking under extrinsic (visual) and intrinsic (BRF) feedback yet rely on distinct biomechanical strategies. Our results indicate that participants increased TLA and ankle moment under visual feedback, whereas they increased hip moment and hip power under BRF, with ankle power adjustments depending on task demand or target level. It would be of interest to perform follow-up studies in larger samples of healthy adults and in post-stroke individuals. These studies should investigate and compare both feedback types during split-belt treadmill walking. Focus should be given to how the paretic limb employs different biomechanical strategies to achieve various propulsion targets.