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

The study investigates whether augmenting residual muscle–tendon afferents enables continuous neural control of a bionic limb to restore biomimetic gait after transtibial amputation. Contemporary powered prostheses often rely on intrinsic state machines or pattern recognition to approximate gait, providing only partial neural control. Standard amputation procedures remove distal tissues and essential proprioceptive afferents, limiting neuromodulation capacity. The authors hypothesize that enhancing residual muscle afferent signaling via an agonist–antagonist myoneural interface (AMI) will allow users to fine-tune supraspinal and spinal circuitry to produce biomimetic gait across varying speeds, terrains, and perturbations without predefined intrinsic robotic controllers.

Prior work on lower-limb prostheses emphasizes intrinsic controllers (finite state machines, pattern recognition) to manage gait phases and terrains. While such systems can improve function, they limit versatility and continuous user-driven modulation. Peripheral nerve stimulation has shown promise in upper-limb neuroprostheses and some lower-limb applications for sensory restoration, improving function, speed, and metabolic cost. However, locomotion depends heavily on spinal reflex circuits and native proprioceptive feedback, and the required magnitude of afferents for high-bandwidth control in gait has been unclear. Earlier AMI research demonstrated improved volitional swing-phase control but did not achieve full-cycle neural control (stance under intrinsic control). The literature indicates that intact individuals use native afferents for rapid adaptation to disturbances, suggesting that augmenting residual afferents could enable biomimetic gait modulation after amputation.

Design: Prospective, non-randomized clinical study within a registered trial (NCT03913273) assessing AMI effectiveness in transtibial amputees. Fourteen unilateral below-knee amputees participated (7 AMI, 7 non-AMI controls (CTL)), matched on age, time since amputation, height, and weight. Inclusion: age 18–65, K3–K4, proficient with passive prostheses, healed residuum, no pain affecting gait. Exclusions included cardiopulmonary instability, pregnancy, and active smoking. All provided informed consent (MIT IRB protocol 1812634918). Recruitment/data collection: April 2019–January 2023 at MIT. Neuroprosthetic interface and system: AMI surgically links residual agonist–antagonist muscle pairs (lateral gastrocnemius–tibialis anterior for ankle; peroneus longus–tibialis posterior for subtalar) to recreate natural muscle dynamics and leverage muscle spindle/Golgi tendon organ proprioception. The autonomous bionic limb comprised a powered ankle (max 162 Nm DF–PF torque; ROM 10° DF/20° PF; 2.42 kg plus 0.33 kg battery), flexible surface EMG electrodes (TA and GAS), and a portable EMG unit (2 kHz, 5 channels, 57 g). Electrodes routed under liner socks and socket; calibration performed once at session start. Control strategy: Continuous neuromodulation without intrinsic state machines or pattern recognition. EMG from TA/GAS processed (band-pass 90–330 Hz, artifact rejection; RMS 200 ms window) to compute activation; a decoding pipeline mapped muscle activity to target equilibrium ankle angle (θref) and an impedance modulation level (μ2). Torque command derived via impedance control using δ = θref − actual angle, μ2, and measured joint state (angle, velocity), bounded by empirically derived biological force–length and force–velocity relationships and passive tissue contributions. Additional virtual damping and stiffness ensured stability at zero muscle activity. Residual afferent estimation: During maximum phantom dorsiflexion and plantarflexion cycles, simultaneous ultrasound measured muscle fascicle strains of residual TA/GAS and surface EMG recorded activations. A computational type II muscle spindle model estimated afferent firing (Imp s−1). Agonist–antagonist afferent bandwidth was defined as the difference between antagonistic (lengthening) and agonistic (shortening) afferents across the TA–GAS pair. Ten cycles per direction; measurements normalized by cycle. Ambulatory testing: Following two practice sessions (~6 h total exposure), subjects performed: (1) level-ground 10-m walks at slow speed (~1.0 m s−1), AMI maximum self-selected speed, CTL maximum speed, and an AMI moderate speed matched to CTL maximum; (2) terrain adaptations: 5° slope decline/incline and staircase descent/ascent; (3) perturbed walking: obstacle crossing (0.21 m high sponge) with subsequent recovery step, in a subset (n=10; 6 CTL, 4 AMI). Safety rails were available; participants encouraged to reduce reliance as comfortable. Data collection and metrics: Onboard sensors and goniometers recorded ankle kinematics/torque and bilateral lower-extremity joint kinematics (1 kHz). Ankle power and net work computed and normalized to body mass; 15 gait cycles for level walking, 10 for each terrain condition. Event-based kinematic metrics included stance push-off and swing foot clearance for level/slope; weight acceptance and forward continuance/foot placement (stair descent); pull-up and foot clearance (stair ascent). LEK symmetry computed from symmetry index across ankle, knee, hip. Perturbation metrics: increase in swing DF vs. unperturbed walking, recovery step peak power and net work, and perturbed walking speed. Torque–angle loops analyzed for energetics. Statistics: Normality via Shapiro–Wilk (α=0.05); repeated-measures ANOVA for AMI speed effects; two-sided paired and unpaired t-tests (Holm–Šidák corrections) as appropriate; Mann–Whitney U-tests when non-normal. Correlations between bionic metrics and agonist–antagonist afferents via Pearson r with 95% CI and slopes. PCA used to summarize gross functionality (PC1) across gait conditions.

• Residual muscle afferents: AMI significantly augmented TA–GAS agonist–antagonist afferents (10.5 ± 0.88 Imp s−1) versus CTL (0.09 ± 0.69 Imp s−1), n=7 per cohort, unpaired t-test, P = 7.7 × 10−7; approximately 18% of intact values (~60 Imp s−1).
• Level-ground walking: AMI subjects exhibited more biomimetic ankle kinematics across speeds, with distinct push-off and swing foot clearance (n=7 per cohort, unpaired t-tests, P < 0.037). AMI increased ankle peak power and net work with speed (n=7, paired t-tests, P < 0.049 and 0.0042), whereas CTL showed no significant changes (P = 0.073 and 0.11).
• Maximum speed and energetics: AMI maximum walking speed was 41% higher than CTL (1.78 ± 0.04 vs 1.26 ± 0.07 m s−1; n=7 per cohort; P = 5.4 × 10−10). AMI peak power was 187% higher (1.95 ± 0.11 vs 0.68 ± 0.17 W kg−1; P = 3.1 × 10−5), and net work increased 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 (3.0 W kg−1) and net work (0.267 J kg−1). AMI maximum speed matched fast-walking speeds of intact individuals (1.78 ± 0.04 vs 1.81 ± 0.03 m s−1; P = 0.52).
• Symmetry: AMI improved lower-extremity kinematic symmetry by 19–26% across speeds compared with CTL (n=7 per cohort; P < 6.2 × 10−10). At maximum speed, AMI LEK symmetry (84.9 ± 1.3%) achieved 93% of intact (91.4%).
• Terrain adaptation: AMI demonstrated biomimetic slope and stair adaptations without changing control parameters. On slopes, AMI maintained level-walking mechanics and increased peak power and net work on incline vs decline (n=7; P = 0.0037 and 0.0045); CTL showed no significant adaptation (P = 0.37 and 0.34). On stairs, AMI modulated shock absorption (descent) and propulsive torque (ascent), with higher positive/negative peak power magnitudes than CTL by 163–202% (n=7 per cohort; P < 0.0052). CTL exhibited restricted ankle ROM and limited adaptability.
• Perturbation response (subset n=10): AMI increased swing DF during obstacle crossing versus unperturbed walking by 4.14 ± 0.66°, while CTL change was negligible (between-group P = 0.0052). During the recovery step, AMI produced greater propulsion: peak power 1.01 ± 0.09 W kg−1 and net work 0.125 ± 0.022 J kg−1 vs CTL 0.45 ± 0.20 W kg−1 and −0.016 ± 0.018 J kg−1 (P = 0.038 and 0.0011). AMI traversed the perturbation 54% faster (0.94 ± 0.05 vs 0.61 ± 0.06 m s−1; P = 0.0049).
• Afferent–function correlations: Residual muscle afferent magnitude strongly correlated with maximum speed (r = 0.85, P = 2.3 × 10−6), ankle peak power (r = 0.88, P = 3.2 × 10−7), net work (r = 0.78, P = 0.0011), LEK symmetry (r = 0.83, P = 1.3 × 10−6), and with biomimetic slope/stair adaptations (|r| = 0.56–0.87, P < 0.035). PCA PC1 (gross functionality) correlated with afferents for unperturbed conditions (r = 0.92, P = 3.6 × 10−8) and perturbed walking (r = 0.90, P = 4.4 × 10−8). In perturbation trials, afferents correlated with increased swing DF, recovery-step propulsion, and perturbed speed (r = 0.64–0.81, P < 0.046).

The findings support the hypothesis that augmenting residual muscle–tendon afferents enables continuous, human-driven neuromodulation of a bionic ankle, restoring biomimetic gait mechanics, energetics, symmetry, and adaptability across speeds, terrains, and perturbations without reliance on intrinsic robotic control. The AMI recreates agonist–antagonist dynamics that engage native sensory organs, providing proprioceptive feedback that tunes both supraspinal and spinal circuits. Despite residual muscles not directly loading against the environment and afferent magnitudes being modest relative to intact (~18%), users achieved substantial functional restoration, indicating strong human sensorimotor adaptability. The observed torque–angle loops progressed from dissipative to propulsive as afferents increased, mirroring intact gait energetics. The strong correlations between afferent bandwidth and multiple functional metrics suggest a dose–response relationship, offering a design principle: even partial restoration of proprioceptive signaling can markedly enhance neuromodulated control. Compared with intrinsic controllers, full neural control leverages the nervous system’s versatility, potentially improving embodiment and clinical outcomes. Future integration of additional sensory modalities (cutaneous, osseoperception) may further enhance performance.

This study demonstrates that continuous neural control of a powered ankle-foot prosthesis, enabled by augmented residual muscle afferents via the AMI interface, restores biomimetic gait after transtibial amputation. AMI users achieved higher maximum walking speeds, increased ankle power and net work, improved kinematic symmetry, natural terrain adaptations, and rapid perturbation responses. Functional improvements scaled with the magnitude of residual muscle afferents, indicating that partial proprioceptive restoration suffices to unlock high-bandwidth neuromodulation. Future work should: (1) evaluate long-term clinical outcomes and embodiment compared with passive and intrinsically controlled devices; (2) extend to multi-DOF joints and other amputation levels; (3) incorporate additional afferent channels (cutaneous feedback, osseointegration) and assess performance in higher-bandwidth tasks (sprinting, jumping, balance) and dynamic perturbations; and (4) refine surgical and interface designs to maximize afferent restoration.

• Study design: prospective but non-randomized and unblinded; small sample size (n=14; perturbed walking subset n=10). Matching was used, but residual confounders may remain.
• Scope: Single powered ankle DOF evaluated; stance interactions in the residuum do not perfectly replicate intact muscle–tendon loading. Only a static perturbation (obstacle crossing) was tested; broader, more dynamic conditions were not assessed.
• Short-term exposure: Approximately 6 hours of practice; long-term adaptation, durability, and daily-life outcomes were not measured.
• Afferent magnitude: Augmented afferents remained well below intact levels (~18%), and ultrasound/EMG-based spindle modeling provides estimates rather than direct neural recordings.
• Generalizability: Participants were relatively high-functioning (K3–K4) transtibial amputees; results may not generalize to other levels or functional statuses.