Regenix — Knee Rehab Exoskeleton

An ultra-low-cost semi-rigid knee exoskeleton

Anthony Wu with the Regenix prototype mounted on the test mannequin, showing the hip actuator pack, cable, five-link knee chain and shin brace

Regenix — A Low-Cost Knee Rehabilitation Exoskeleton

Regenix addresses the lack of affordable knee rehabilitation in low- and middle-income settings, where supervised physiotherapy is out of reach and robotic exoskeletons cost USD 7,000–100,000. It is an ultra-low-cost, semi-rigid, cable-driven knee exoskeleton with a bill of materials under USD 200, combining a self-adaptive five-link chain with a distributed load-anchoring system. Its hardware includes an ESP32 microcontroller, dual IMUs, a cable-driven actuator, and 3D-printed parts, weighing under 2 kg. The controller runs four rehabilitation modes across 0°–145° of motion and delivers at least 10 N·m of assistive torque. Mannequin testing showed ≤1° inter-trial leg-angle deviation, and a kinematic model predicts 1.39 mm mean joint misalignment. An on-device analytics framework combining K-Means clustering, trend analysis, and LLM recommendations reached 94% agreement with licensed clinicians. As the first sub-USD-200 knee exoskeleton with integrated on-device AI, Regenix enables at-home rehabilitation in settings without physiotherapists.

Regenix Project Video

Reel · 2 min walkthrough

Business & Commercialization

I worked with my friend Leo Xie to develop Regenix’s business model, conducting market research, competitor analysis, and customer segmentation. We built financial projections targeting a $250 retail price, analyzed manufacturing costs, and explored subscription-based revenue models. I researched FDA medical-device regulations, outlining a potential Class II 510(k) clearance pathway, clinical validation requirements, and international regulatory considerations. We also analyzed healthcare reimbursement policies, insurance coverage, and socioeconomic barriers to rehabilitation access, integrating these findings into our commercialization and market-entry strategy.

Anthony Wu presenting the Regenix prototype and rehabilitation dashboard at the Diamond Challenge
Anthony Wu explaining the Regenix research poster and prototype to a judge at the science fair
Anthony Wu and Leo Xie at the Diamond Challenge

Publication

Awards

Diamond Challenge

Semi-finalist
Diamond Challenge

MSEF

2nd Place · Region V
Massachusetts Science & Engineering Fair

Conrad Challenge

Innovator Award
Conrad Challenge

University of Rochester

Xerox Award for Innovation & Information Technology
University of Rochester

Project notes
01
01 — Problem

Commercial powered exoskeletons — ReWalk, Ekso, Roam — retail at US $7,000 to $100,000, an order of magnitude above the annual per-capita health budget of every low- or middle-income country on Earth.

In the United States alone, an estimated 38 million people with chronic musculoskeletal conditions do not access outpatient physical therapy due to systemic barriers.

At the same time, the two cheap alternatives — rigid post-op braces and soft compression sleeves — fail biomechanically: the rigid hinge fights the knee's migrating center of rotation, and the soft sleeve cannot deliver therapeutic torque.

FIG · COST LADDER · USD $0 $25k $50k $75k $100k $90,000 $70,000 $7,000 $200 — under 3% of the cheapest powered competitor. ReWalk 6.0 FULL LOWER LIMB Ekso GT GAIT TRAINER Roam Ascend CONSUMER KNEE EXO Regenix THIS WORK
Fig. 1.2 Per-unit retail price. Regenix's $200 bar is so small it nearly disappears on this axis — that is the point.
FIG · RIGID-EXOSKELETON MISALIGNMENT 0° → 90° FLEXION a RIGID EXOSKELETON · FIXED HINGE CONSTRAINT PROXIMAL CUFF shear force at thigh strap FIXED HINGE PIVOT axis cannot migrate with ICR DISTAL CUFF shear force at shin strap KEY ELEMENTS FIXED HINGE PIVOT single rigid axis — cannot follow ICR migration CUFF SHEAR ZONES constraint forces generate skin friction under load The hinge is fixed. The joint is not. b ANATOMICAL VIEW · KNEE ICR FRAME FEMUR proximal segment θ · FLEXION ANGLE between bone long-axes rICR · MIGRATING ORIGIN local frame (x, y) TIBIA distal segment c THE MISALIGNMENT METRIC M(θ) = ‖ rICR(θ) − rexo(θ) ‖ M(θ) ↑ ⇒ Fmisalignment ↑ As the knee flexes, rICR migrates along the J-curve while rexo stays pinned. Every degree of flexion grows M — and with it the shear force on the cuff.
Fig. 1.1 Rigid exoskeletons assume the knee rotates about a single fixed axis. It doesn't — the instantaneous center of rotation (ICR) traces a J-shaped curve through flexion, so every fixed pin generates shear at the cuff and contact forces at the joint.
02
02 — Device

A segmented chain wraps the posterior knee — flexible when the joint moves freely, rigid when the Dyneema cable tensions through it — guiding force along a path that tracks the knee's own instantaneous center of rotation.

A single hip-mounted brushed-DC actuator does all the work; the limb itself carries only aluminium rails, PETG cuffs, and a chain that field-repairs with 3D-printed spares.

Component
Semi-rigid chain
Segments9 × PETG links TendonDyneema · ⌀ 1.0 mm
Loading chain… dragorbit
Fig. 2.1 Posterior chain assembly. Discrete PETG links thread onto a single Dyneema tendon — slack when the leg swings, rigid the instant the cable is tensioned, wrapping the joint along its true center of rotation.
The Regenix prototype mounted on a life-size anthropomorphic mannequin, showing the hip-mounted actuator pack, the Dyneema cable routing, and the posterior chain running to the lateral knee cuff.
Fig. 2.2 Bench-test deployment on an anthropomorphic mannequin. Hip pack carries the DH03X actuator, ESP32, and battery; a single Dyneema tendon routes down the lateral thigh into the chain and terminates at the shank cuff — distal mass under 250 g.
FIG · ALIGNMENT COMPARISON 0° → 145° ROM
FIXED HINGE
Constrains the knee.
Δ misalign θ ≈ 0° EXTENSION shear ↑ θ ≈ 60° FLEXION
SEMI-RIGID CHAIN
Tracks the joint.
Loading…
motor
@ hip
← tracks ICR
stable l_a
θ ≈ 0°  chain straight
θ ≈ 60°  chain wraps
FIG · ALIGNMENT FIXED HINGE Constrains the knee. Δ misalign θ ≈ 0° EXTENSION shear ↑ θ ≈ 60° FLEXION SEMI-RIGID CHAIN Tracks the joint. θ ≈ 0° CHAIN STRAIGHT l_a stable θ ≈ 60° CHAIN WRAPS
On-board sensing & comms
Three breakout boards — under $17 combined — handle force, motion, and the clinician hand-off.
HX711 ADC and S-type load cell
HX711 + load cell
Cable tension · ±5 kg
MPU-6050 6-axis IMU breakout
MPU-6050
6-axis IMU · knee angle
DX-BT18 Bluetooth serial module
DX-BT18
Bluetooth · clinician link
03
03 — How it works

The user puts on the device and powers it on — the twelve-link semi-rigid chain passively conforms to their individual knee geometry. From there, they activate one of three rehabilitation modes using the on-board voice control: "Hey Regenix, start Assistive training."

Regenix mirrors the three stages of rehabilitation. Each mode trades the device's authority for the patient's own, in proportion to how much strength they have recovered.

Stage 01
Assistive
Motor leads.
The motor helps the patient complete movements when strength is still limited — full-arc flexion and extension on cue.
DevicePatient
Stage 02
Protective
Motor follows.
Assistance only activates after the user generates some force themselves — encouraging early muscle engagement without overload.
DevicePatient
Stage 03
Resistive
Motor opposes.
The device adds controlled resistance to rebuild strength as recovery progresses — closed-loop torque against the patient's own effort.
DevicePatient

After each session, the training data feeds into an on-device AI pipeline that analyzes it and generates a full rehabilitation report — flagging the user's current training stage, projected recovery timeline, and quantitative metrics that doctors can act on.

01
Voice command
"Hey Regenix, start Assistive."
→
02
Guided session
Mode-specific torque profile, 8–15 min.
→
03
On-device analysis
Range, torque, symmetry, adherence.
→
04
Clinician report
Stage, timeline, recommended next session.
Fig. 3.1 Closed-loop session pipeline. The device runs the same four steps every session — the report from step 4 informs the mode for step 1 of the next.
04
04 — Result

On a life-size anthropomorphic mannequin, the prototype tracked the knee's center of rotation with a mean misalignment of 1.39 mm across the full 0°–145° range, delivered 13.5 Nm peak torque, and reproduced peak cable forces with a coefficient of variation under 2 %.

An on-device analytics pipeline — statistics → trend → K-Means → composite score → LLM — generated rehabilitation recommendations that agreed with licensed physiotherapists in 82 % of sessions, without ever sending data to the cloud.

$200
BOM target
0.85kg
Total mass
13.5Nm
Peak torque
1.39mm
Mean misalign.
82%
AI · PT agreement
145deg
Range of motion

Validation was performed on a hinged anthropomorphic mannequin under six experimental conditions (initial leg angle ∈ {75°, 95°}, servo speed ∈ {9, 13.5, 20 °/s}). Inter-trial leg-angle deviation ≤ 1° across three repeatability runs; force RMSE ≤ 0.8 N against a calibrated reference load.

Continue to the deep dive.

Explore the theoretical, mechanical and software design along with the experimental data and what comes next

Open the full paper →
Interactive prototype
Regenix Model
Knee θ
72.0°
Cable F
180 N
Torque τ
8.40 Nm
Loading specimen… dragorbit ⇧pan ⌫⌫reset