Why Bedside Physical Therapy Robots Need Force-Controlled Joint Modules？

A therapy robot does not sell a robot arm; it sells a force applied to a human body. Pressure, rhythm and depth are the prescription, and position control cannot deliver them: soft tissue is compliant and time-varying, so one position command yields contact forces that differ several-fold between muscle and bone. Force-controlled harmonic joint modules turn contact force into a closed-loop variable - the only way to make dose repeatable, human contact safe and a therapist's technique teachable. Starting from one site photo of a bedside therapy robot, this article reads the configuration, sets out five reasons force control is unavoidable, settles HONPINE's TCHL-versus-HAG choice criterion by criterion, and gives the torque sensor resolution budget that picks the frame sizes.

1. Reading the Machine: What Kind of Robot Is This?

One site photograph already tells you most of what matters for selection. The machine is a 6-7 DOF vertical articulated arm working over a therapy bed, built from large-diameter one-piece joint modules. Nothing here is high dynamic: the duty cycle is low speed, sustained contact, long sessions. What the photo does show - external corrugated conduit running across several joints - is a real integration pain point, and it is the argument for a large hollow shaft. All readings below are visual estimates used to demonstrate a method, not the actual design parameters of that machine.

2. Five Reasons Force Control Is Non-Negotiable in Therapy Robots

2.1 The stiffness paradox: position control does not govern force on compliant tissue

Position control assumes a rigid, compliant-free environment. Human soft tissue is the opposite: low stiffness, non-linear and time-varying. The same position command meets thick muscle and a bony prominence and produces contact forces that differ by a factor of several; breathing and how far the patient relaxes shift the force again at an unchanged position. Contact force under position control equals position error x environment stiffness x arm equivalent stiffness - and not one of those three terms is machine-governed.

Force control changes the controlled variable from position to force, so contact force becomes a closed-loop quantity. That is the only control architecture that can take responsibility for the force applied to a human body.

2.2 Therapy dosimetry: force is the prescription, and it must be repeatable

The clinical parameters of manual therapy are force (or pressure), frequency, duration and depth. The commercial promise of a therapy robot is standardised force and repeatable treatment - the same prescription should feel the same on the first session and the tenth, with this therapist or that one, in this clinic or that one. Without a force loop, "dose" is only a number in software, not a physical fact.

2.3 Human-contact safety: collision detection and safe stop are the compliance floor

Patients move, cough and turn over without warning. Torque-sensing joints give the whole arm collision detection through joint torque residual monitoring, and pair with STO (safe torque off) to turn an abnormal contact event into a safe torque-off stop. That is the compliance basis for entering any human-contact scenario: ISO 10218 and ISO/TS 15066 provide the collaborative reference, while a device classified as a medical device must be worked through IEC 60601 and ISO 14971 risk management. Certification conclusions rest with the notified body, not with this article.

2.4 Hand-guided teaching: the therapist's touch has to be transferable

Today the technique lives in the hands of experienced therapists, so the product route is fixed: the therapist guides the arm through the movement, the machine records trajectory and force, then replays it. Hand guiding requires the joints to run in zero-force or compliant mode - a direct application of torque closed loop. Without force control, guiding is either immovable or collides at the slightest push.

2.5 Compensation and unit-to-unit consistency

Harmonic joints carry friction and hysteresis, and every change of arm posture changes the gravity torque. Measured joint torque is the data source for gravity and friction feed-forward; compensate badly and the mechanical side pollutes tip force accuracy. Equally, the calibration record shipped with each module - stiffness and hysteresis curve - is what keeps "every robot feels the same" true across a batch.

3. Two Layers, Not Either-Or: Joint Torque Sensing and End-Effector Force Sensing

The most common specification error is to treat joint torque sensing and a wrist force/torque sensor as alternatives. They are two layers with different jobs, and a therapy robot needs both. The joint layer measures joint torque, not contact force: it has to be mapped through link length and posture, and it is contaminated by gravity, inertia and harmonic friction, so reading it directly as "massage force" carries real error. The end-effector layer sees only contact beyond the wrist - it cannot detect an arm-body collision and cannot serve joint safety or compensation.

4. TCHL vs HAG: Item-by-Item Verdict for Therapy Robots

HONPINE's TCHL and HAG are two product definitions on one platform, not two generations. TCHL is the lightweight, compact series built for overseas channels; HAG is the highly integrated series with functional safety. Both offer an optional torque sensor, and that is where the verdict turns.

4.1 Where HAG wins outright

4.1.1 A published torque sensor specification you can promise

A force-controlled product has to answer one question in writing: what is the resolution? With HAG the answer is published - range per frame and 0.5% of full scale non-linearity - so a resolution figure can be quoted, calibrated and defended. TCHL publishes "optional" only. On a product whose entire value proposition is force, an unspecified sensor is a hard defect.

4.1.2 Integral STO as the functional-safety base

Physical human contact needs a safety floor that does not depend on software. HAG integrates STO from frame 17 upward, which is what turns collision detection into a safe torque-off. Confirm the safety level and certification status per frame in writing before you commit.

4.1.3 Multi-turn absolute encoder for patient positioning

After an e-stop or a power loss in a clinical session, the patient is still on the bed. A single-turn absolute encoder forces a re-homing routine with a human being under the arm; HAG's 12-bit multi-turn removes that step entirely.

4.2 Where TCHL still makes sense

TCHL remains the better answer where mass is the binding constraint - wearable or airborne structures, portable devices carried by an operator - and where the host controller speaks CAN FD. On a bedside therapy robot with a fixed or wheeled base, weight is not the binding constraint, so its main advantage simply does not apply.

5. Frame Matching and the Force-Control Resolution Budget

5.1 Published HAG sensor ranges and resolution

Resolution is range x 0.5% of full scale, because non-linearity is quoted against full scale. The practical rule follows directly: choose the range that sits close to the working torque. Oversizing costs resolution - a 500 N.m sensor resolves 2.5 N.m, twenty times coarser than a 25 N.m sensor.

5.1.1 Worked example: from tip force to joint torque

A manual therapy normal force of 10-50 N, taken through a wrist moment arm of roughly 0.10-0.20 m, becomes a 1-10 N.m wrist torque signal. Against the table above, HAG-17's 0.375 N.m resolution is about 2-4 N at the tip - workable but not generous for a prescription graded in 5 N steps, and the joint reading still carries gravity, inertia, and harmonic friction and hysteresis on top.

5.2 The boundary that has to be stated: joint sensing is not therapy-dose closed loop

A joint-level torque sensor does not by itself close the loop on therapy dose. The correct architecture is two layers: joint torque sensing for safety, collision detection and compensation, and end-effector force sensing for dose. Saying this out loud is what separates a credible specification from a marketing claim - and it is also the honest reason to keep a wrist F/T sensor in the bill of materials.

6. Getting HAG Right: Ordering and Validation Checklist

6.1 Before you order

1. STO per frame. Confirm the frame number in the model code HAG-XX T-XXX-DC1-02-B-E actually carries STO, and ask for the safety level (SIL/PL) and certification status. The HONPINE documentation contradicts itself on STO (HAU-14 "not included" versus HAG-14 "integrated"), so this has to be clarified in writing.

2. Measurement state of published figures. HAG ships in T (with sensor) and N (without) configurations, but the catalogue carries a single parameter table. Get written confirmation of whether rated and peak torque were measured with or without the torque sensor.

3. Sensor dynamic behaviour. Bandwidth, hysteresis and thermal drift. Therapy runs continuously for long sessions: drift decides whether minute 30 matches minute 1, and bandwidth caps the achievable compliance loop.

4. How torque readings reach the bus, and how often. Decide whether the force loop runs inside the drive or on the host. Under 1 kHz control the sensor reporting rate has to match - prefer the EtherCAT version.

6.2 Before you ship

5. Factory calibration record. Require a stiffness, hysteresis and zero-drift calibration report per module, and confirm the controller can read those parameters for compensation. This is the handle on "every machine feels the same" in volume delivery.

6. Bench validation items. Tip force step response and steady-state accuracy at room temperature and after warm-up; collision detection threshold and response time together with STO; hand-guiding feel; accuracy decay after one million cycles.

7. Compliance pathway. Where the device falls under medical device rules, run risk management against IEC 60601 and ISO 14971. This article is not a certification basis; the professional body's opinion governs.

7. One-Page Conclusion

Why force control. What a therapy robot delivers is a force on a human body, and a human body is a deformable contact object, so position control cannot govern force. Force control is the precondition, not the premium option: dose, safety and teachable technique all sit on it.

TCHL or HAG. HAG. Three decisive advantages land exactly on the pressure points of this scenario: a force specification that can be promised (published range and 0.5% non-linearity), an STO functional-safety base, and a multi-turn absolute encoder that recovers patient positioning. TCHL's lightweight advantage carries the least weight in a bedside machine, and an unpublished sensor specification is a hard defect on a force-control product.

How to use it. Take wrist HAG-17, elbow HAG-20 and shoulder HAG-25 as the baseline combination and check range margin joint by joint against working torque. Run two layers - joint torque sensing for safety and compliance, end-effector force sensing for dose - and close the seven items in section 6 before releasing an order.

Sources and confidence levels

Disclaimer

· Configuration readings (degrees of freedom, joint diameter, reach, torque tiers) are visual estimates, not the real design parameters of the machine photographed. They illustrate a selection method and must not be presented as a technical conclusion about that model.

· Manufacturer figures come from catalogues that contain several internal inconsistencies; this article does not silently pick one reading. Before any external commitment, work from the latest specification and a written confirmation from the manufacturer.

· Where a therapy device is classified as a medical device, safety and compliance (IEC 60601 series, ISO 14971 and related) rest with the professional certification body. This article is not a certification basis.