Every servo joint can report torque two ways: estimated from motor current, or measured by a strain-gauge sensor on the output flange. The first is free: the current loop already exists. The second costs money and a few millimetres of length. The difference between them is everything the reducer does to the torque on its way through.
Why current lies, and by how much
Motor torque is genuinely proportional to current; that part is physics and it's reliable. The trouble is that your process feels output torque, and between motor and output sit the strain wave gear's losses: tooth friction, flexspline hysteresis, grease drag. Current-based estimation must subtract a friction model from the measurement, and that model is wrong in three well-known ways:
- Stiction. At zero and near-zero speed — exactly where delicate contact happens — friction is discontinuous. The estimator can't tell a gentle touch from breakaway friction.
- Temperature. Grease drag changes several-fold between a cold start and a warm afternoon. A friction model calibrated at 25 °C reads fiction at 5 °C.
- Age. Friction rises as the gear wears: slowly, monotonically, uncalibrated.
The practical floor: through a high-ratio harmonic drive, current-based output-torque estimates are typically uncertain by a few percent of rated torque at best, and far worse near zero speed. For a joint rated at tens of newton-metres, that's whole newton-metres of fog exactly where fine tasks live.
What a real sensor changes
An output-side torque sensor measures after the reducer: friction, hysteresis, and temperature drift all happen upstream of it. We build two sensed families: the AS-SF force-sensing joints (±35 to ±220 N·m, accuracy ≤0.5% of full scale, 30 Hz closed force-loop bandwidth) and the torque-sensing AS-SJ3 servo joint frames, which extend the sensed range from ±25 up to ±500 N·m. Concretely: on an AS-SF-70 (±35 N·m), 0.5% F.S. is about ±0.18 N·m of measurement truth, an order of magnitude below where current estimation is still guessing.
When current is enough
Plenty of force problems don't need the truth, just a bound:
- Collision detection on a fenced industrial axis: a threshold at, say, 20% of rated torque trips fine on current alone.
- Torque limiting to protect a mechanism: clamp the current, accept the friction margin.
- Heavy contact tasks where the working force is a large fraction of rated torque: a few percent of fog on a big number is noise.
The pattern: when the force you care about is large compared to the joint's friction, current-based control is honest work at zero cost.
When you need the sensor
- Fine assembly and insertion. The signal of interest (fractions of a newton-metre at the tool) is below the friction fog. No model rescues it.
- Surface-following. Polishing, deburring, ultrasound scanning: constant normal force through changing pose, where gravity compensation and friction both vary with geometry.
- Physical human interaction. Hand-guiding and power-and-force-limited collaborative operation want low, verified thresholds with sensing that doesn't drift with temperature or age. Standards work in this space (ISO 10218 / ISO/TS 15066-class limits) is much easier to demonstrate against a calibrated sensor than a friction model.
- Process verification. When the torque applied is the product record (fastening, dispensing pressure), an estimate isn't evidence.
The decision, compressed
Ask one question first: is the force you must control smaller than the joint's own friction, referred to the output? If yes, specify a force-sensing joint; no firmware will out-clever the physics. If no, start with current-based limits, and leave yourself the upgrade path: the AS-SF frames are drop-in siblings of the standard joints, so the decision can be revised per axis without redesigning the arm.
Borderline case? Describe the task and the force band. Telling you honestly whether you need the sensor is exactly the kind of question the RFQ form exists for.
Related: Choosing a robot-joint reducer · Harmonic drive efficiency.