BLOGS

In brief: An integrated joint is usually specified by torque, ratio and mass — numbers that describe the parts, not the joint. Once the motor, reducer, drive, encoders and brake share one housing, the binding constraints move to three other layers: how deterministically the control loops run, what the feedback actually measures, and how heat and stiffness couple through a shared structure. A joint drive-and-control platform disclosed in September 2026 by a chip developer working with a Chinese Academy of Sciences institute makes the first layer measurable, and the numbers are far more instructive than the gear ratio.

The Joint's Three Ceilings

Every actuator has a transmission ceiling, a feedback ceiling and a compute ceiling. The joint performs at whichever is lowest.

The transmission ceiling is mechanical. A harmonic (strain wave) reducer delivers near-zero backlash and a high single-stage ratio of roughly 50:1 to 160:1 inside a very short axial envelope, because it transmits torque through the elastic deformation of a thin-walled flexspline. That same deformation is the limit: the output lags the input under load, and the lag is not constant — it changes with torque direction, temperature and wear. Planetary stages trade backlash for stiffness and shock tolerance. Cycloidal (RV) stages buy the highest torsional rigidity at a volume and mass premium. No transmission family escapes the trade.

The feedback ceiling is what you can observe. If the only encoder sits on the motor shaft, the controller knows how far the rotor turned and nothing about what happened after the gearbox.

The compute ceiling is how fast, and how predictably, the controller can act on what it observes. This is the layer that has quietly become the bottleneck — and the one that a joint's specification sheet almost never describes.

Why a Single Core Is Not Enough

Field-oriented control (FOC) needs microsecond-class loops: sample phase currents, run the transform math, update the PWM, repeat. A fieldbus stack, meanwhile, cannot afford to drop a frame. Run both workloads on one core and they compete for the same cycles, which means determinism and stability become hard guarantees rather than design properties.

The engineering conclusion reached by the team behind the September 2026 disclosure is worth stating plainly: the bottleneck in joint control usually sits in the compute architecture, not in the control algorithm. Their response was to split the workloads across a heterogeneous dual-core device — one core dedicated to FOC motor control, the torque/speed/position three-loop scheme, dual-encoder data processing and safety functions, the other dedicated to CAN FD, EtherCAT and serial communication, with the two exchanging commands and status through on-chip SRAM and hardware semaphores.

The peripheral budget on the motor-control side is not decorative: three 12-bit ADCs sampling at 5 Msps for phase-current and bus-voltage sensing, ultra-high-resolution timers for complementary PWM generation, on-chip flash and SRAM sized to hold the control algorithms and protocol stacks together, and an operating range from −40 °C to +105 °C with ECC-protected RAM and windowed watchdogs. Integrating all of that into the joint shortens the analogue signal chain, which is exactly where noise and latency are cheapest to remove.

The Numbers You Should Be Asking For

Bench results from the same disclosed platform (rated torque 20 N·m, peak torque 60 N·m) give a useful checklist of the metrics that actually describe joint behaviour. Ask any integrated-joint vendor for these before peak torque.

The position figure is the one worth dwelling on. Roughly half a degree of static error is the difference between grasping a soft object without crushing it and not — and that error is produced by the whole chain, not by the encoder alone.

What Dual Encoders Actually Fix

Put a second absolute encoder on the gearbox output and the transmission ceiling stops being invisible. The error sources it exposes are the ones a motor-side encoder structurally cannot see:

• reducer backlash, where reversing direction makes the output lag the command
• elastic deformation, where the motor turns and the output does not follow
• a small mismatch between the nominal and actual reduction ratio
• drift from temperature and wear over the service life

With load-side feedback, four capabilities open up. Position control can close on the true joint angle instead of a derived one. The live difference between the two encoders becomes a transmission-error model that can pre-compensate the position target under load. In a compliant joint, that same difference multiplied by the known drivetrain stiffness yields an estimate of joint torque without a dedicated torque sensor. And a divergence between the two encoders that does not match the model becomes a fault signal for slippage, breakage or collision.

Typical resolutions in current humanoid-grade joints run around 17 bits or better on the motor side, where high speed matters, and 19 bits or higher on the output side, where absolute multi-turn accuracy matters. Dual 19-bit absolute configurations are now routine in humanoid joints rather than exotic.

There is one trap that lives outside the datasheet: magnetic encoders are disturbed by magnetic fields. Brakes and cables routed through a hollow shaft are both offenders, so critical joints often need field-guiding geometry or shielding, and some builders move to inductive encoders on the axes that cannot tolerate the interference.

Not Every "Dual Encoder" Is a Dual Encoder

A joint can carry two sensors and still be blind on the output. Three configurations are commonly marketed as dual-encoder designs without providing load-side feedback: one physical encoder with the output angle derived in software from the ratio; two encoders both mounted on the motor side, such as an incremental encoder plus a Hall sensor; and an output-side sensor with resolution so coarse that it can only serve as a reference point or limit switch.

The test is simple. Is there an independent, high-resolution position sensor on the load output whose signal participates in closed-loop control or error compensation? If the answer is that the output angle is computed from the motor-side reading, the transmission error is still unmeasured — and unmeasured error cannot be compensated.

Wire the Two Loops in Series, Never in Parallel

Once both feedback paths exist, the control structure matters as much as the sensors. Averaging the two readings is a common and costly mistake: the motor-side encoder is fast and high-bandwidth but measures the rotor, while the output-side encoder measures what the customer cares about but carries the drivetrain's lag. They are not interchangeable estimates of the same quantity.

The workable arrangement is cascaded. The current and velocity loops close on the motor-side encoder, giving the fast inner loop the responsiveness it needs. The position loop closes on the output-side encoder, using the velocity loop as its inner loop. A compliance compensation term is then added to the position target to absorb reducer deformation and backlash before the loop ever sees them. In short: the joint-side encoder decides where the joint should be, and the motor-side encoder decides how to get there.

The Coupling Nobody Puts on the Datasheet

Integrating the drive into the joint housing is not free. Signal paths get shorter and EMI behaviour improves dramatically when encoder signals run on a dedicated shielded layer or board-level differential pair — but the drive's switching losses now conduct straight into the motor and the reducer.

Continuous torque is therefore limited by heat, not by winding capability, and thermal design becomes the structural design problem. Two further couplings are worth naming. Stiffness first: transmission stiffness and structural stiffness together set the joint's resonant frequency, and the closed-loop bandwidth can never exceed it. Teams routinely fail to raise bandwidth not because the controller is mistuned but because the stiffness chain ran out of margin first. And routing second: the hollow bore that lets cables and air lines pass through the axis is now reaching the 20 mm class, which changes wall thicknesses, bearing selection and sealing strategy at the same time.

Add protection and qualification to the list. IP54 to IP67 sealing, grease behaviour across the operating temperature range, and bearing seal design are hard gates for industrial dust and outdoor duty. Before a prototype is machined, static strength, fatigue, modal, thermal-magnetic-structural and EMC analysis should already have filtered out the predictable failure modes.

How to Evaluate an Integrated Joint

1. Ask for loop rate and determinism, not just loop rate. A 20 kHz current loop driven by an architecture that stalls under bus traffic is not a 20 kHz joint.
2. Verify the encoder topology physically. Confirm there is a real absolute sensor on the output, and that its signal reaches the position loop.
3. Request loaded accuracy, not unloaded accuracy. Backlash and elastic deformation only appear once torque is applied.
4. Treat peak torque as a transient rating. The continuous rating and the thermal path decide whether the joint survives a shift.
5. Check the stiffness chain against your bandwidth target. If your controller needs 200 Hz and the structure resonates at 150 Hz, no tuning will save it.
6. Confirm bus compatibility at project kickoff. EtherCAT, CAN FD or serial decided late becomes weeks of integration rework.
7. Audit EMI on the bench, not in the field. Run the brake and the hollow-shaft cabling while watching encoder noise.
8. Buy integration, not parts. Motor, reducer, drive and dual encoders tuned as one unit remove the mismatch that component-level selection quietly introduces.

The Ti5ROBOT View

We build joint modules around exactly this stack. Our Humanoid Robot Actuator line integrates the frameless motor, drive electronics, feedback and mechanical structure as one unit, and our Lightweight Joint Module series applies the same integration discipline to compact, precision-critical axes where envelope and mass are the binding constraints.

The reasoning above is the electrical counterpart of a theme we have explored on the mechanical side: our joint map of harmonic, planetary and RV drives sorts transmissions by joint family, and our review of what torque sensing changed for a force-controlled arm shows what the sensing layer unlocks once feedback and control are designed together. A joint is not the sum of its parts; it is the weakest of its three ceilings.

The Bottom Line

Specify an integrated joint and the datasheet will talk about torque, ratio and mass. The behaviour you actually get is decided elsewhere: in whether control loops have dedicated deterministic compute, in whether the output angle is measured or inferred, and in how heat and stiffness propagate through a shared structure. The reducer still sets the mechanical envelope. But in 2026 the accuracy ceiling is usually written in the electronics.

Source: Ti5ROBOT Blog, September 2026. Based on a joint drive-and-control platform disclosed in September 2026 by a chip developer collaborating with a Chinese Academy of Sciences research institute, together with published engineering guidance on dual-encoder joint architectures, reducer selection and integrated joint thermal design. Figures are quoted as single-source benchmark disclosures unless corroborated by multiple sources; no pricing information is included by design.