Continuous neural control of a bionic limb restores biomimetic gait after amputation

Contemporary bionic legs typically rely on intrinsic controllers (for example, finite state machines and pattern recognition) that replay predefined gait patterns with limited user neuromodulation. After standard amputations, loss of distal tissues and peripheral afferents impairs reflexive and volitional control, hampering biomimetic gait when using neural inputs alone. Prior work in upper-limb systems shows that artificial sensory feedback can improve control, but locomotion depends heavily on spinal reflex circuitry and may require richer, more biomimetic afferents. The research question here is whether augmenting residual muscle afferents in transtibial amputees enables continuous, high-bandwidth neural control of a bionic leg to restore biomimetic gait mechanics, adaptability across terrains, and robust perturbation responses without intrinsic robotic gait controllers. The authors hypothesize that even modest augmentation of residual muscle afferents, achieved via an agonist–antagonist myoneural interface (AMI) and continuous EMG-based control, will enable biomimetic gait neuromodulation.

• Current lower-limb prostheses often use intrinsic control architectures (finite state machines, pattern recognition) with limited neural input, yielding partial user control constrained to specific gait phases or terrains.
• Standard lower-limb amputation removes distal tissues and essential proprioceptive afferents, degrading reflexive spinal contributions to gait.
• Peripheral nerve stimulation has enhanced sensation and control mainly in upper-extremity neuroprostheses, with some lower-extremity benefits, but fully biomimetic, fully neurally driven gait has not been demonstrated.
• Intact human locomotion rapidly adapts to disturbances, suggesting that augmenting residual afferents could allow amputees to fine-tune neural circuitry for biomimetic gait.
• Prior AMI work showed improved volitional free-space control and swing-phase kinematics but did not achieve fully neurally driven stance-phase control or quantify how afferent levels relate to biomimetic gait restoration.

Study design: Prospective, non-randomized clinical study within registered trial NCT03913273. Fourteen unilateral transtibial amputees (age 47.6 ± 3.5 years; time since amputation 3.9 ± 0.5 years) participated: 7 had AMI amputations (AMI cohort) and 7 had standard/non-AMI amputations (CTL cohort), matched by age, time since amputation, height, and weight. Inclusion: proficient with passive prostheses, variable cadence (K3–K4), healed limb, no gait-limiting pain. Exclusion: cardiopulmonary instability, pregnancy, active smoking. IRB approvals and informed consent obtained.

Neuroprosthetic interface and system: The AMI surgically links residual agonist–antagonist muscle pairs to recreate natural dynamics (lateral gastrocnemius to tibialis anterior for ankle; peroneus longus to tibialis posterior for subtalar). Surface flexible EMG electrodes recorded residual tibialis anterior (TA) and gastrocnemius (GAS). A powered ankle-foot prosthesis (162 Nm max DF–PF torque, 10° DF/20° PF ROM, 2.75 kg system) was controlled by a continuous EMG-driven impedance controller. The controller decoded target equilibrium angle (θref) and impedance modulation (µ2) from TA/GAS activations and constrained torque by angle- and velocity-dependent biological torque limits and passive structure estimates. No intrinsic gait phase detection, state machines, or pattern recognition were used. Subject-specific decoding was calibrated once via maximal phantom DF/PF while standing.

Residual muscle afferent estimation: During 10 cycles of maximum phantom DF and PF, simultaneous ultrasound measured fascicle strain and surface EMG recorded TA and GAS activity. A type II muscle spindle model estimated afferent firing (Imp s−1). Agonist–antagonist muscle afferents were computed as the difference between antagonistic (lengthening) and agonistic (shortening) afferents for the TA–GAS pair during steady-state portions (25–75% of cycle). Group comparisons used t-tests or Mann–Whitney U-tests as appropriate.

Gait testing: After two practice sessions (~6 hours total), subjects performed:

• Level-ground walking in a 10 m hallway at: slow (target ~1.0 m s−1 for all), self-selected maximum speed (both cohorts), and for AMI also a moderate speed matched to CTL maximum (~1.25 m s−1).
• Terrain adaptation on a 5° slope (decline/incline) and stairs (descent/ascent) at self-selected speeds, with handrails allowed if needed.
• Perturbed walking (subset n=10: CTL 1–6, AMI 1–4): obstacle crossing using a 0.21 m-high sponge; recovery step analyzed.

Data acquisition and metrics: Onboard sensors and goniometers recorded bionic ankle angle, torque, and ipsilateral/contralateral hip, knee, ankle kinematics at 1 kHz. Derived metrics included ankle power and net work (normalized to body mass), stance/swing event-based kinematics (push-off, foot clearance; stair weight acceptance and pull-up/foot placement), range of motion, and lower-extremity kinematic (LEK) symmetry (100 − symmetry index). Fifteen gait cycles per level-ground condition and 10 per terrain condition were collected. For perturbations, swing DF change vs. unperturbed baselines, recovery step peak power/net work, and perturbed walking speed were computed. Torque–angle loops characterized gait energetics.

Statistical analysis: Normality by Shapiro–Wilk. Within-AMI repeated-measures ANOVA for speed effects; paired and unpaired two-sided t-tests with Holm–Šidák corrections; Mann–Whitney U-tests for non-normal data. Pearson correlations (r), 95% CI, slopes (m), and P values assessed relationships between agonist–antagonist afferents and functional metrics across conditions. PCA summarized gross functionality (PC1 scores) for unperturbed and perturbed conditions.

• Residual muscle afferents: AMI significantly augmented agonist–antagonist muscle afferents vs CTL (10.5 ± 0.88 vs 0.09 ± 0.69 Imp s−1; n=7 per cohort; unpaired t-test, P=7.7×10−7), corresponding to ~18% of intact values (~60 Imp s−1).
• Level-ground walking energetics and speed: At maximum self-selected speeds, AMI achieved 41% higher speed than CTL (1.78 ± 0.04 vs 1.26 ± 0.07 m s−1; n=7 per cohort; P=5.4×10−10), with 187% higher peak power (1.95 ± 0.11 vs 0.68 ± 0.17 W kg−1; P=3.1×10−5) and higher net work by 0.223 J kg−1 (0.174 ± 0.015 vs −0.049 ± 0.054 J kg−1; P=0.0018). AMI restored ~65% of intact peak power and net work (intact ~3.0 W kg−1 and 0.267 J kg−1), and reached intact fast-walking speed (1.78 ± 0.04 vs intact 1.81 ± 0.03 m s−1; P=0.52).
• Speed adaptation: AMI increased bionic ankle peak power and net work with walking speed (P<0.049 and 0.0042), mimicking intact speed-dependent reflex contributions; CTL showed no significant changes (P=0.073 and 0.11).
• Kinematics and symmetry: AMI neuromodulated distinct push-off and foot-clearance angles across speeds (P<0.037) and improved LEK symmetry by 19–26% across speeds vs CTL (n=7 per cohort; P<6.2×10−10). At maximum speed, AMI LEK symmetry was 84.9 ± 1.3%, 93% of intact (91.4%).
• Terrain adaptation: Without parameter changes, AMI exhibited biomimetic slope and stair adaptations, maintaining level-ground mechanics while modulating energetics (incline vs decline peak power and net work; P=0.0037 and 0.0045). AMI produced greater terrain-specific positive/negative peak power (163–202% higher than CTL; P<0.0052). CTL failed to show biomimetic adaptations and had restricted ankle ROM on slope and stairs (significant between-group differences reported).
• Perturbation response (subset n=10): During obstacle crossing, AMI increased swing DF vs unperturbed by 4.14 ± 0.66°, whereas CTL showed negligible change (−0.45 ± 0.87°; P=0.0052). In the recovery step, AMI generated higher propulsion: peak power 1.01 ± 0.09 vs 0.45 ± 0.20 W kg−1 (P=0.038) and net work 0.125 ± 0.022 vs −0.016 ± 0.018 J kg−1 (P=0.0011), enabling 54% faster perturbed walking (0.94 ± 0.05 vs 0.61 ± 0.06 m s−1; P=0.0049).
• Dose–response with afferents: Increasing residual afferents transformed torque–angle loops from clockwise (dissipative) to counterclockwise (positive work) and strongly correlated with peak power (r=0.88), net work (r=0.78), maximum speed (r=0.85), and LEK symmetry (r=0.83) at maximum speed (n=14; P≤0.0011). Terrain and perturbation metrics similarly correlated with afferents (|r|≈0.56–0.90; P<0.046). PCA PC1 scores of gross functionality correlated strongly with afferents for unperturbed (r=0.92; n=14; P=3.6×10−6) and perturbed walking (r=0.90; n=10; P=4.4×10−4).
• Overall, modest augmentation (~18% of intact) of residual muscle afferents enabled continuous neuromodulation of biomimetic, versatile gait across speeds, terrains, and perturbations without intrinsic robotic gait controllers.

The study demonstrates that augmenting residual proprioceptive muscle–tendon afferents via AMI reconstruction enables people with transtibial amputation to continuously neuromodulate a powered ankle and restore biomimetic gait mechanics. Compared to intrinsic controller paradigms that discretize intent, the full neural control scheme leveraged the user’s nervous system to adapt across gait phases, speeds, terrains, and perturbations. Despite total afferent magnitudes being far below intact values, participants exhibited robust improvements in energetics, kinematics, symmetry, terrain versatility, and obstacle negotiation, implying substantial human sensorimotor adaptability. The strong dose–response between afferent bandwidth and functional metrics supports the central role of proprioceptive feedback in lower-limb locomotion and provides design principles for surgical reconstruction and neural interfaces. Clinically, such neuroprostheses may reduce secondary complications and improve embodiment by restoring more natural control; adding complementary afferents (cutaneous feedback, osseointegration) may further enhance functionality and embodiment.

Augmenting residual muscle afferents with an AMI-based neuroprosthetic interface and employing continuous EMG-driven control (without intrinsic robotic gating) restores biomimetic gait in transtibial amputees. Even modest afferent augmentation (~18% of intact) enabled neuromodulation of speed-dependent energetics, symmetric multi-joint kinematics, terrain adaptability (slope, stairs), and robust perturbation responses, achieving near-intact fast walking speeds. The clear correlations between afferent levels and functional outcomes provide actionable targets for surgical and interface design. Future research should: evaluate long-term clinical benefits and embodiment versus passive/intrinsic powered prostheses; expand testing to higher-bandwidth tasks (sprinting, jumping, balance) and dynamic perturbations; incorporate additional sensory channels (cutaneous, osseoperception) and multi-DOF control; and generalize to other amputation levels and neuroprosthetic applications.

• Study design was prospective, non-randomized, with small cohorts (n=7 per group) and a subset for perturbation testing (n=10); participants and researchers were not blinded.
• Short-term laboratory evaluation; long-term clinical outcomes (metabolic cost, pain, musculoskeletal health, embodiment) were not assessed.
• Only transtibial amputees were studied; generalizability to other levels and multi-DOF joints remains to be shown.
• Perturbation testing addressed a static obstacle only; responses to dynamic/irregular terrains and balance perturbations were not evaluated.
• Despite functional gains, residual afferent magnitudes remained far below intact levels; the specific contributions of different afferent modalities (muscle vs cutaneous vs osseoperception) were not isolated.
• Initial user perception of device heaviness suggests acclimation effects; training duration and learning dynamics were limited (~6 hours).