Regenix — Knee Rehab Exoskeleton

Design Process — failure-driven iteration, from a bench rig to a wearable, repeatable system

Could a cable-driven chain assist a knee at all?

The first stage answered one question before any refinement: can a hip-mounted motor, a cable and a chain behind the knee deliver useful assistance? Each build exposed the failure the next one had to fix — cable routing first, then force transmission — until the load path held on a life-size mannequin.

Bench-top conceptual rig: aluminium frame, motor, weight and chain
01Conceptual design

A bench rig on an aluminium frame: motor, spool, weight and chain, used to confirm that tensioning a cable through a segmented chain produces knee-extension torque.

First wearable prototype on a person, cable routed down the leg
02Inconsistent cable routing

The first wearable build. The tendon path shifted from trial to trial, so the same motor command produced different assistance at the joint.

Second wearable prototype with soft anchoring on the shank
03Instable force transmission

Routing was fixed, but the anchors were not: cuffs migrated under load and the force never reached the knee cleanly or repeatably.

Final prototype on the mannequin: semi-rigid chain, shin brace and distal anchor
04Final product

Semi-rigid five-link chain on rigid thigh and shin anchors. A stable load path on the mannequin — the baseline every Stage 2 iteration was measured against.

Sensing, control and the cable spool

With the concept proven, iteration turned to the two subsystems that decide whether assistance is accurate: the electronics that sense and command the joint, and the spool that turns motor rotation into cable tension. Both were refined by failure.

Iterative development of the sensing and control architectureIntegration of multimodal sensing and an optimized system architecture enabled reliable closed-loop control aligned with real-time biomechanical feedback.

Version 1 wiring: Arduino Uno with HX711 load cell and BT18 Bluetooth module
v1 · Arduino Uno + HX711 + BT18
  • Basic force measurement and Bluetooth communication.
  • Lacked motion sensing and closed-loop control, limiting its ability to respond to user biomechanics.
Version 2 wiring: Arduino Uno with HX711, dual MPU6050 IMUs and DH-03X motor
v2 · + dual MPU6050 + DH-03X
  • Dual MPU6050 IMUs introduced to capture thigh–shank kinematics.
  • Integration with the DH-03X motor allowed preliminary closed-loop control.
  • Lacked proper wiring infrastructure and space to integrate a signal-transmission module.
Version 3 wiring: ESP32 with motor driver, dual MPU6050 IMUs and Wi-Fi signalling
v3 · ESP32 integrated system
  • Integrates an ESP32 Wi-Fi signalling module with motor driver and both IMUs on one board.
  • Enables real-time data acquisition and closed-loop control.

Failure-driven refinement of the cable spool and routing mechanism

Spool v1: bare pulley on the motor shaft
Spool 1
  • Uncontrolled cable crossover during winding led to frequent jamming.
  • The varying winding radius produced inconsistent torque output and reduced mechanical efficiency.
Spool v2: dual-channel spool
Spool 2
  • A dual-channel geometry was introduced to separate cable paths and reduce interference.
  • Insufficient constraint allowed radial drift and jamming of the cable between the two channels.
Spool v3: red 3D-printed spool with cable anchor and partial routing constraints
Spool 3
  • Improved cable anchoring and partial routing constraints stabilized the tendon path and reduced slippage.
  • Torque transfer became more consistent, but minor variations in effective radius still affected precision and efficiency.
Spool v4: fully constrained winding mechanism with fixed cable path and optimized spool geometry
Spool 4 · final
  • A fully constrained winding mechanism with fixed cable paths and optimized spool geometry.
  • Ensures a constant effective radius of 0.033 m, stable tension delivery and repeatable torque aligned with the knee's flexion-dependent biomechanics.

Structure, routing and clean force sensing

The last round of iteration fixed what the wearable itself did to the force: a frame that broke, then one that was too heavy; a cable path that bled torque into friction; and a load cell that read every spike. Each fix was driven by a measured failure.

Failure-driven design evolution of the exoskeleton structure and cable routing

June prototype: thin 3D-printed shank structure with red load-cell housing
JuneBroke in use

Inadequate structural connection and cable instability, causing component breakage during operation.

July prototype: reinforced white brace with thicker chain mount
JulyStrong but heavy

Addressed the strength limitations but introduced excessive mass and geometric inefficiency.

December prototype: compact black brace with guided cable routing and integrated chain mount
DecemberStable and wearable

Incorporates optimized structural geometry and guided cable routing to ensure stable force transmission while maintaining compactness and wearability.

Optimization of cable routing to reduce frictional losses

Cable routed between the legs through a red guide, with a sharp bend near the actuator
Before

Cable routed between the legs and required a sharp ~90° directional change near the actuator — high friction from bending and surface contact.

Cable director beside the actuator and its CAD part
After

Cable rerouted along the waist through a compliant rubber interface, with a cable director beside the actuator to keep alignment and prevent misrouting — reduced friction and more efficient, stable force transmission.

Optimization of load-cell integration using buffering

Load cell rigidly coupled in line with the cable on the red housing
Before

The load cell was rigidly coupled to the cable, so it directly experienced transient force spikes during actuation — high signal noise and poor repeatability.

Compliant spring in series with the load cell on the shin anchor
After

A compliant spring element in series with the load cell absorbs transient spikes and smooths the applied load — mechanical filtering that reduces high-frequency noise.

What the iteration produced

Four things Regenix does differently from the exoskeletons it was measured against: how it aligns with the joint, where it puts the load, what it does with the data, and how it trains the patient.

Innovation 1 · Low-cost semi-rigid joint alignment mechanism

Forces around a rigid knee joint: motors, Bowden cable sheath, bracket, anchor and load cell
Traditional

A rigid hinge fixes one axis; the knee's own axis migrates through flexion, so every degree adds misalignment and cuff shear.

Fig. 4 — Liu et al., 2022
Semi-rigid chain structure: straight and flexed link drawings
Innovation

Adaptive chain eliminates axis mismatch while maintaining a deterministic tendon moment arm.

Fig. 5 — Semi-rigid chain structure

Innovation 2 · Distributed load anchoring system

Joint torques and reaction forces on a rigid knee hinge exoskeleton
Traditional

Peak tension reacts straight through the knee hinge and its cuffs — the joint surface carries the load.

Fig. — Witte et al., 2017
Placement of waist belt, hip attachment, strap A and anterior-thigh anchor
Innovation

Multi-strap routing distributes peak tension across waist belt, hip and anterior thigh — zero contact on the knee surface.

Fig. 6 — Placement of exoskeleton anchors

Innovation 3 · Embedded lightweight AI analytics

Traditional

Exoskeletons are only able to provide assistive torque and record measurements.

Innovation

K-Means clustering + statistical analysis + an LLM API, running on low-cost microcontrollers, interpret the results on-device.

Innovation 4 · Interactive rehabilitation modes

1 · Assistive training mode

Provides active torque assistance to guide the knee through its range of motion, supporting early-stage rehabilitation when patients cannot generate sufficient force independently.

Initial position → position after active assistance
2 · Protective training mode

Delivers assistance only after a user-generated force threshold is detected, ensuring safe, effort-initiated movement while preventing under- or overexertion.

Active leg extension → force-change detection → active assistance
3 · Resistive training mode

Applies adaptive resistance proportional to user input and joint motion, promoting muscle strengthening and neuromuscular control during mid-to-late rehabilitation.

Active hamstring curl → force-change detection → resistance (cable rotates backwards)

What it is built from

The final assembly and the electronics inside it: a hip-mounted actuator pack, a Dyneema extension-assist cable, the five-link semi-rigid knee chain and a shin anchor, controlled by an ESP32 with IMU and load-cell sensing.

Overall wearable assembly: waist actuator pack with motor, spool and control; extension-assist cable; thigh cuff load path; five-link semi-rigid knee chain; shin brace and distal anchor
Overall wearable assemblyWaist actuator pack → extension-assist cable → thigh cuff load path → five-link semi-rigid knee chain → shin anchor.

Electronics used

ESP32
ESP32Control & connectivity
DC actuator
DC actuatorHip-mounted drive
HX711 + load cell
HX711 + load cellCable tension sensing
MPU-6050
MPU-6050Knee-angle sensing
DX-BT18
DX-BT18Bluetooth link