Sim-to-Real Transfer and Human-in-the-Loop Optimization for Personalized Lower-Limb Assistance

Reinforcement Learning-Adaptive Rehabilitation Exoskeletons for Post-Stroke Gait Recovery

1. Problem Statement

Stroke is the leading cause of long-term disability in the United States. The CDC reports more than 795,000 new strokes annually, with someone experiencing a stroke every 40 seconds. Of the approximately 7.6 million stroke survivors in the US, an estimated 50-60% retain chronic lower-limb motor deficits affecting gait, balance, and independent mobility (Virani et al., "Heart Disease and Stroke Statistics — 2024 Update," Circulation, 2024). This translates to approximately 3.8-4.6 million Americans living with stroke-related gait impairment.

The current standard of care for post-stroke gait rehabilitation is physical therapist-supervised overground and treadmill training, typically delivered in 60-minute sessions, 3-5 times per week, over 4-12 weeks. The American Heart Association/American Stroke Association guidelines recommend a minimum of 3 hours per week of task-specific gait training. In practice, therapist availability and cost ($150-300 per session) constrain the dose most patients receive. A 2024 meta-analysis of 34 randomized controlled trials (1,166 participants) confirmed that robotic exoskeleton-assisted gait training significantly improves walking velocity, step length, cadence, and functional independence compared to conventional therapy alone (Frontiers in Neurology, 2024).

The economic burden is substantial. The CDC estimates annual stroke-related medical costs exceed $56 billion in the United States, encompassing acute care, rehabilitation, and lost productivity. Inpatient rehabilitation alone costs $17,000-35,000 per stroke patient (Agency for Healthcare Research and Quality). The therapist-to-patient ratio is the primary throughput bottleneck: each rehabilitation therapist can work with only one patient at a time during gait training, and physical therapist vacancy rates exceed 20% in many rural and underserved rehabilitation facilities.

Existing rehabilitation exoskeletons — ReWalk (FDA-cleared, personal use HCPCS K1007, Medicare reimbursement $94,617), Ekso Bionics EksoNR (institutional), and Indego (Parker Hannifin) — deliver fixed, pre-programmed gait trajectories. These devices follow the same kinematic pattern regardless of the patient's current motor capacity, fatigue level, or recovery trajectory. A patient who improves from 20% to 60% of normal gait function receives the same assistance profile at both stages. The therapist manually adjusts parameters between sessions, but the device does not adapt within a session or across sessions autonomously. This fixed-trajectory paradigm limits rehabilitation effectiveness because motor learning theory — specifically the challenge point framework (Guadagnoli and Lee, Motor Control, 2004) — demonstrates that optimal learning occurs when assistance is calibrated to the learner's current skill level, neither too easy nor too difficult.

The unmet market need is a rehabilitation exoskeleton that uses reinforcement learning to continuously adapt its assistance to the individual patient's neurological state, reducing therapist burden while optimizing motor recovery through personalized, progressive challenge.

2. State of the Art

Four parallel developments have converged to define a clear commercial opportunity in RL-adaptive rehabilitation exoskeletons.

Human-in-the-loop optimization has quantified the metabolic benefit of personalized assistance. Zhang et al. (Science, 2017) at Carnegie Mellon demonstrated that CMA-ES optimization of ankle exoskeleton torque patterns, tuned to individual users during walking, reduced metabolic energy consumption by 24.2 ± 7.4% — the largest reduction demonstrated at that time. Slade et al. (Nature, 2022) at Stanford extended this approach outdoors using Bayesian optimization, achieving 23 ± 8% metabolic reduction during treadmill walking and a 9 ± 4% increase in self-selected walking speed, with optimization converging four times faster than laboratory methods. These results establish that personalized assistance outperforms fixed parameters by a wide margin.

Sim-to-real reinforcement learning has eliminated the need for per-user experiments. Luo et al. (Nature, 2024) at North Carolina State University trained a universal RL controller entirely in simulation using a 50-degree-of-freedom musculoskeletal model with 208 skeletal muscles, then deployed it directly on a custom 3.2 kg bilateral hip exoskeleton without any human experiments or parameter tuning. The system reduced metabolic cost by 24.3% for walking, 13.1% for running, and 15.4% for stair climbing (n=8 per activity). Training required only 8 hours on a single NVIDIA RTX 3090 GPU. This is the critical advance: it demonstrates that RL controllers can be designed entirely in simulation and transferred to physical hardware, eliminating the expensive and slow per-user calibration process.

Task-agnostic control has generalized across activities. Molinaro et al. (Nature, 2024) at Georgia Institute of Technology demonstrated a deep neural network controller that estimated biological hip and knee joint moments in real-time (R² = 0.83 vs. ground truth) and provided task-agnostic assistance across 28 activities — from level walking and running to lunging, lateral cutting, and unstructured meandering. The controller reduced user energetics by 5.3-19.7% across 10 tested activities without any manual adjustment between tasks. The exoskeleton was integrated into athletic pants rather than a bulky robotic frame.

RL-based adaptive control under pathological conditions is emerging. Chavarrias et al. (arXiv, 2025) demonstrated an RL-based adaptive torque controller for a knee exoskeleton that adjusts assistance under spasticity — the involuntary muscle stiffness that affects 20-35% of stroke survivors. Using a digital twin with a differentiable spastic reflexes model, the TD3 agent reduced maximum joint torques by 10.6% and decreased tracking error settling time by 8.9% compared to conventional compliant control. Current commercial exoskeletons cannot accommodate spasticity; patients scoring above 1+ on the Modified Ashworth Scale are excluded from exoskeleton rehabilitation. An adaptive controller that handles spasticity would expand the eligible patient population by 20-35%.

All four research directions have been validated on human subjects or in validated computational models. None has been combined into a single rehabilitation exoskeleton product. All FDA-cleared rehabilitation exoskeletons continue to use fixed gait trajectories with manual therapist adjustment.

3. Foundational Research

Zhang J, Fiers P, Witte KA, Jackson RW, Poggensee KL, Atkeson CG, Collins SH. "Human-in-the-loop optimization of exoskeleton assistance during walking." Science, 356(6344):1280-1284, 2017. DOI: 10.1126/science.aal5054. PMID: 28642437. The Carnegie Mellon/Nanyang Technological University team applied covariance matrix adaptation evolution strategy (CMA-ES) to optimize ankle exoskeleton torque profiles for individual users during treadmill walking. The optimizer varied four parameters controlling torque onset timing, peak magnitude, and duration across approximately 32 minutes of walking per participant. Optimized profiles reduced metabolic energy consumption by 24.2 ± 7.4% compared to zero-torque walking. This result is foundational because it quantified, for the first time, the gap between generic and personalized exoskeleton assistance — a 24% metabolic benefit from tuning four parameters in a single joint. A multi-joint adaptive system optimizing dozens of parameters across the full gait cycle would be expected to achieve substantially larger benefits for impaired walkers.

Slade P, Kochenderfer MJ, Delp SL, Collins SH. "Personalizing exoskeleton assistance while walking in the real world." Nature, 610(7931):277-282, 2022. DOI: 10.1038/s41586-022-05191-1. PMID: 36224415. Stanford researchers developed a data-driven method for outdoor exoskeleton optimization using only wearable sensors — no laboratory instrumentation required. Bayesian optimization personalized ankle exoskeleton parameters during one hour of naturalistic walking in a public setting, converging four times faster than laboratory CMA-ES methods. The optimized assistance increased self-selected walking speed by 9 ± 4% and reduced metabolic energy per unit distance by 17 ± 5% outdoors. On a treadmill at 1.5 m/s, metabolic reduction reached 23 ± 8%. The significance for rehabilitation is direct: this demonstrates that personalization can occur during normal walking, not just in a clinical gait laboratory, enabling continuous adaptation in home and community settings.

Luo S, Jiang M, Zhang S, Zhu J, Yu S, Dominguez Silva I, Wang T, Rouse E, Zhou B, Yuk H, Zhou X, Su H. "Experiment-free exoskeleton assistance via learning in simulation." Nature, 630(8016):353-359, 2024. DOI: 10.1038/s41586-024-07382-4. PMID: 38867127. The NC State/University of Michigan team trained an RL policy entirely in simulation using a comprehensive musculoskeletal model (50 degrees of freedom, 208 muscles), then deployed it on a custom 3.2 kg bilateral hip exoskeleton with zero human experiments. The system used two nine-axis IMU sensors on the thighs and ran inference on a Raspberry Pi 4. Metabolic reductions: 24.3% walking, 13.1% running, 15.4% stair climbing (n=8 per activity). Training required 8 hours on a single GPU. This is the enabling advance for clinical scalability: sim-to-real transfer eliminates the need for per-patient optimization sessions, replacing hours of therapist-supervised calibration with a simulation-trained policy that works immediately on any new user.

Molinaro DD, Scherpereel KL, Schonhaut EB, Evangelopoulos G, Shepherd MK, Young AJ. "Task-agnostic exoskeleton control via biological joint moment estimation." Nature, 635(8038):337-344, 2024. DOI: 10.1038/s41586-024-08157-7. PMID: 39537888. Georgia Tech researchers trained a deep neural network to estimate biological hip and knee joint moments from wearable sensors in real-time, then used these estimates to drive proportional exoskeleton assistance. The network achieved R² = 0.83 for joint moment estimation across 28 diverse activities. When tested on 10 activities (level walking, running, lifting 25 lb, lunging, stair climbing, and others), the controller reduced metabolic cost or biological joint work by 5.3-19.7% with zero manual tuning between activities. The exoskeleton was integrated into clothing-like athletic pants. For rehabilitation, this demonstrates that a single controller can handle the variability of real-world movement — patients do not walk on flat treadmills exclusively, and a clinically useful exoskeleton must accommodate sit-to-stand transfers, stair negotiation, and outdoor terrain.

Chavarrias A, Sanz-Morere CB, De Rossi C. "Adaptive Torque Control of Exoskeletons under Spasticity Conditions via Reinforcement Learning." arXiv:2503.11433, 2025. Spanish National Research Council researchers developed a TD3 reinforcement learning agent to control a knee exoskeleton under spasticity conditions — the first RL controller explicitly designed for pathological muscle tone. Using a digital twin combining a musculoskeletal-exoskeleton model with a differentiable spastic reflexes model, the agent reduced maximum joint torques by 10.6% and decreased root mean square error settling time by 8.9% compared to conventional compliant control. Spasticity affects 20-35% of stroke survivors and is the primary exclusion criterion for current exoskeleton rehabilitation; this controller addresses the clinical population most underserved by existing devices.

Frontiers in Neurology systematic review (2024). "Effect of robotic exoskeleton training on lower limb function, activity and participation in stroke patients: a systematic review and meta-analysis of randomized controlled trials." Meta-analysis of 34 randomized controlled trials involving 1,166 participants demonstrated that lower-limb robotic exoskeleton training (LRET) significantly improved motor control, walking velocity, step length, cadence, and functional independence compared to conventional rehabilitation alone. This meta-analysis establishes Level 1 evidence that exoskeleton-assisted gait training is effective — the question is no longer whether exoskeletons work for stroke rehabilitation, but how to optimize their assistance for individual patients.

4. Competitive Landscape

Ekso Bionics (San Rafael, CA). Market leader with 17% global exoskeleton market share. Products include EksoNR (institutional rehabilitation) and Ekso Indego Personal. FDA-cleared for stroke and spinal cord injury rehabilitation. Annual revenue approximately $38 million (2024). All devices use pre-programmed gait trajectories with manual therapist adjustment via tablet interface. No published work on RL-adaptive control. In December 2025, Ekso announced a proposed merger with Applied Digital Corporation, which could result in separation or sale of its exoskeleton business — signaling potential strategic uncertainty in their rehabilitation product line.

Lifeward/ReWalk Robotics (Yokneam, Israel). ReWalk 7 launched in the US in April 2025, offering personal exoskeleton use for spinal cord injury with customizable walking speeds and app connectivity. Medicare reimbursement established at $94,617 per device (HCPCS K1007, CMS 2024 fee schedule). Total historical funding approximately $150 million. Partnership with AlterG bundles anti-gravity treadmill therapy with exoskeleton training. All devices use fixed gait patterns. No adaptive RL control.

Indego (Parker Hannifin, Cleveland, OH). Indego Personal exoskeleton for SCI. FDA-cleared. Distributed through orthotics and prosthetics providers. Fixed trajectory control with therapist-selectable programs.

No commercial rehabilitation exoskeleton uses reinforcement learning, human-in-the-loop optimization, sim-to-real transfer, or any form of adaptive closed-loop control that personalizes assistance based on the patient's real-time physiological or biomechanical state. All three market leaders use the same control paradigm: pre-defined kinematic trajectories with manual parameter adjustment. The competitive gap is not in the hardware — all existing devices have adequate actuators, sensors, and structural design — but in the control software. The team that demonstrates RL-adaptive control on existing or comparable hardware and navigates the FDA pathway for an AI-controlled rehabilitation device captures a software-defined market advantage in a $579 million exoskeleton market growing at 16.25% CAGR (Precedence Research, 2025).

5. Total Addressable Market

Bottom-up calculation (US, post-stroke gait rehabilitation — initial indication):

Annual new strokes in US: 795,000 (CDC, Stroke Facts 2024)
Survivors with chronic lower-limb gait deficits: 50-60% = 397,500-477,000 patients per year
Eligible for exoskeleton rehabilitation (ambulatory, medically stable, Modified Ashworth Scale ≤3 with adaptive control): 60% = 238,500-286,200 patients
Average reimbursable course of exoskeleton-assisted rehabilitation: 24 sessions x $300/session (CPT 97116 gait training + 97530 therapeutic activities, facility rate) = $7,200 per patient
Institutional exoskeleton device revenue: 5,000 rehabilitation facilities x 1.5 devices x $150,000/device (amortized over 5 years) = $225 million device market
Annual rehabilitation services TAM: 260,000 patients x $7,200 = $1.87 billion
Institutional device TAM: $225 million (replacement cycle)
Personal exoskeleton market: 50,000 qualifying patients/year x $94,617 (Medicare K1007 rate) = $4.73 billion potential
Conservative US TAM (services + institutional devices): $2.1 billion annually
Full US TAM including personal devices: $3.2 billion annually

Expansion to adjacent indications: Spinal cord injury (296,000 US patients, National SCI Statistical Center), traumatic brain injury (1.5 million/year, CDC), multiple sclerosis (900,000 US patients, National MS Society), and cerebral palsy (764,000 US individuals). The adaptive RL control advantage applies to all conditions requiring gait rehabilitation, extending the total addressable population significantly.

Top-down cross-check:

The global rehabilitation robotics market was valued at $1.5 billion in 2024, projected to reach $4.2 billion by 2033 at 12.5% CAGR (Verified Market Reports, 2024). The US medical exoskeleton market was $283.8 million in 2024 (Grand View Research), and the broader US physical therapy market is $26 billion. AI in rehabilitation and assistive technologies was valued at $2.34 billion in 2024, projected to $8.29 billion by 2034 at 13.6% CAGR (InsightAce Analytic, 2024). An RL-adaptive rehabilitation exoskeleton capturing 10-15% of the US rehabilitation robotics segment yields $200-350 million initially, scaling with market growth — consistent with the bottom-up institutional device and services estimate.

Reimbursement pathways already exist. Exoskeleton-assisted gait training is billable under CPT 97116 (gait training, $41-85/15-minute unit, facility-dependent) and CPT 97530 (therapeutic activities). Personal exoskeletons are reimbursed under HCPCS K1007 at $94,617 (CMS 2024 DMEPOS fee schedule). The adaptive RL component does not require a new reimbursement code — it is a software enhancement to an already-reimbursable device class.

6. Research Gap and Commercial Opportunity

Three specific gaps separate published research from a deployable RL-adaptive rehabilitation exoskeleton:

Gap 1: Adaptation for neurological impairment profiles, not just healthy users. The four landmark studies (Zhang 2017, Slade 2022, Luo 2024, Molinaro 2024) optimized assistance for healthy, able-bodied participants. Post-stroke gait is fundamentally different: asymmetric, slow (0.4-0.8 m/s vs. 1.2-1.5 m/s normal), characterized by circumduction, foot drop, and compensatory hip hiking, and complicated by spasticity, proprioceptive deficits, and fatigue that varies hour to hour. Chavarrias et al. (2025) addressed spasticity in simulation but has not validated in stroke patients. The entity that demonstrates sim-to-real RL transfer on stroke patients — replicating the Luo et al. framework with musculoskeletal models incorporating hemiparetic gait — owns the foundational clinical data for this device class. No academic lab has published this result.

Gap 2: Progressive challenge calibration for motor learning. Current research optimizes for metabolic cost reduction — making walking easier. Rehabilitation requires the opposite: progressively withdrawing assistance to force the patient's nervous system to relearn motor patterns (challenge point framework). An RL controller for rehabilitation must optimize for motor recovery, not energy efficiency — a fundamentally different reward function. The reward signal should incorporate gait symmetry improvement, voluntary muscle activation (EMG), and therapist-defined recovery milestones rather than metabolic cost alone. No published RL exoskeleton controller has been designed with a rehabilitation-specific reward function.

Gap 3: Manufacturing at rehabilitation facility volumes with medical device quality systems. Research exoskeletons are custom-built prototypes: hand-wound motors, 3D-printed brackets, research-grade IMU sensors, and Raspberry Pi controllers. A clinical device must be FDA-cleared, manufactured under ISO 13485, built with medical-grade components (IEC 60601 electrical safety), and serviceable in the field. The Luo et al. exoskeleton weighs 3.2 kg and costs an estimated $15,000-25,000 in research components — a clinical product requires DFM to achieve comparable performance at $40,000-60,000 production cost with 50,000-hour MTBF reliability. No academic lab has manufacturing engineering capability, and exoskeleton companies (Ekso, ReWalk) have not invested in RL control software.

Commercial thesis: The research community has proven that RL makes exoskeletons dramatically better (24% metabolic improvement over fixed control). The rehabilitation market has proven that exoskeletons improve patient outcomes (34-RCT meta-analysis). Reimbursement pathways exist (K1007, CPT 97116). The company that bridges the gap — adapting RL control for neurological impairment, designing rehabilitation-specific reward functions, and manufacturing at clinical quality — captures the adaptive rehabilitation exoskeleton market before the incumbents upgrade their fixed-trajectory software.

7. Comparable Funded Projects

NSF CAREER Award CMMI-1944655. PI: Hao Su, North Carolina State University. "Experiment-free exoskeleton assistance via learning in simulation." This award funded the foundational work that led to the Luo et al. (Nature 2024) paper — the sim-to-real RL controller that eliminated the need for per-user experiments. The resulting publication in Nature validates NSF's investment in this research direction and demonstrates that simulation-trained RL controllers can work on physical exoskeletons without modification.

NSF Future of Work Award 2231419. Partially supported research on adaptive exoskeleton control using reinforcement learning for workplace applications. Demonstrates NSF interest in adaptive exoskeleton technology beyond rehabilitation.

NIH Intramural Research. PI: Thomas C. Bulea, NIH Clinical Center. Neurorobotics Research Group. Developed a pediatric knee exoskeleton with real-time adaptive control for overground walking in children with cerebral palsy. Published in JNER. NIH's own research campus is investing in adaptive exoskeleton control for neurological rehabilitation — validating the research direction.

NIH NIDILRR Center Grant (2015-2020, $4.625 million). Five wearable robot projects including adaptive control for neurological rehabilitation. Demonstrates sustained federal funding for wearable rehabilitation robotics.

CMS Medicare Reimbursement Determination (2024). CMS established HCPCS K1007 and a $94,617 purchase fee schedule for personal rehabilitation exoskeletons, creating a direct revenue pathway for adaptive devices. This regulatory action signals government commitment to exoskeleton rehabilitation as a reimbursable treatment modality.

8. Opportunity Assessment

TRL evidence chain: TRL 5 for the integrated RL-adaptive rehabilitation exoskeleton concept. Individual components are at higher TRL: exoskeleton hardware is TRL 9 (FDA-cleared devices in commercial use), RL-adaptive control for healthy users is TRL 5 (demonstrated on human subjects in real-world conditions, Slade et al. 2022; Luo et al. 2024), and sim-to-real transfer is TRL 5 (deployed on physical hardware, Luo et al. 2024). The integration of RL-adaptive control with rehabilitation-specific reward functions for stroke patients has not been demonstrated — placing the combined system at TRL 4-5.

Top 3 technical risks and mitigation:

RL policy generalization across heterogeneous neurological impairments. Stroke survivors present with widely varying motor deficits: some retain voluntary hip flexion but lack ankle dorsiflexion; others have global hemiparesis. An RL policy trained on one impairment profile may perform poorly on another. Mitigation: the Luo et al. simulation framework can be extended with hemiparetic gait models (published musculoskeletal models of post-stroke gait exist, e.g., Knarr et al., Gait & Posture, 2013); combined with the Slade et al. rapid Bayesian personalization (1 hour), a two-stage approach — simulation pre-training + rapid in-clinic personalization — addresses inter-patient variability without requiring per-patient simulation. Risk level: moderate.
Safety of adaptive assistance in patients with impaired balance. A sudden change in exoskeleton torque could cause a fall in a patient with compromised balance and proprioception. Mitigation: safety-constrained RL with hard torque rate-of-change limits implemented in firmware; falls detected via IMU with immediate torque reduction; all adaptive operation conducted with standard rehabilitation safety equipment (parallel bars, gait belt, fall harness). Existing FDA-cleared exoskeletons already operate under equivalent safety protocols. Risk level: low-moderate.
Regulatory pathway for AI-adaptive Class II medical device. The RL controller constitutes AI/ML-based Software as a Medical Device (SaMD). FDA's 2023 guidance on Predetermined Change Control Plans (PCCP) for AI/ML-enabled devices provides a framework for adaptive algorithms, but no rehabilitation exoskeleton has been cleared under this pathway. If the algorithm is locked after training (sim-trained, not adaptive during patient use), it follows the standard 510(k) pathway with existing cleared exoskeletons as predicates. If the algorithm adapts during patient use, a PCCP must be submitted detailing the adaptation boundaries, safety constraints, and performance monitoring protocol. NeuroPace RNS (closed-loop responsive neurostimulation for epilepsy, PMA P100026) serves as regulatory precedent for an adaptive AI-controlled medical device. Risk level: moderate (but well-defined regulatory pathway exists).

Regulatory pathway: 510(k) with ReWalk or Ekso EksoNR as predicate device. The exoskeleton hardware is substantially equivalent to existing cleared devices. The adaptive software component falls under FDA's guidance for AI/ML-based SaMD. A locked algorithm (trained in simulation, deployed without on-device learning) follows the standard 510(k) pathway. An adaptive algorithm (continues learning from patient data) requires a PCCP per FDA's April 2023 guidance. Classification: Class II medical device (21 CFR 890.3480, Powered Exercise Equipment, or 890.3475, Mechanical Chair). Estimated timeline: 6-12 months pre-submission, 12-18 months clinical evaluation, 6-9 months FDA review. Total: 2-3 years to market with locked algorithm; 3-4 years with adaptive PCCP. Regulatory approval creates a 2-3 year competitive moat — the time required for competitors to generate equivalent clinical evidence and navigate their own 510(k) submissions.

9. Team Requirements

Developing a commercially viable RL-adaptive rehabilitation exoskeleton requires three intersecting capabilities:

Biomedical domain expertise with clinical rehabilitation knowledge. Deep understanding of post-stroke gait pathophysiology (upper motor neuron syndrome, spasticity grading via Modified Ashworth Scale, Brunnstrom stages of motor recovery), rehabilitation outcome measures (10-meter walk test, 6-minute walk test, Functional Ambulation Category, Berg Balance Scale), and clinical trial design for rehabilitation interventions. Required for: defining rehabilitation-specific RL reward functions that optimize motor recovery rather than metabolic efficiency, designing clinical validation protocols that rehabilitation-focused FDA reviewers expect, and identifying the patient population stratification that determines commercial market segmentation.

Machine learning and adaptive control systems. Reinforcement learning expertise: sim-to-real transfer for musculoskeletal systems, human-in-the-loop Bayesian optimization, safety-constrained policy optimization, and transfer learning for patient personalization. Specific experience with biomechanical simulation environments and physiological signal processing (EMG, IMU, gait event detection). Required for: building the adaptive controller that transforms fixed-trajectory rehabilitation into personalized, progressive-challenge-calibrated motor learning — the software that creates the entire competitive advantage.

Manufacturing engineering for wearable medical devices. Design for manufacturability of actuated lower-limb orthotic devices: motor selection and integration, compliant mechanism design for variable-stiffness joints, sensor suite integration (IMU, force/torque, EMG electrodes), battery and power management for 4-8 hour clinical shift operation, and biocompatible skin-contact materials. Quality systems: ISO 13485, IEC 60601-1 (medical electrical equipment safety), design controls per 21 CFR 820. Required for: bridging the gap between research exoskeletons (3D-printed, hand-assembled, $15,000-25,000 BOM) and clinical devices manufactured at 1,000-10,000 units/year with 50,000-hour MTBF and field serviceability.