A humanoid joint is often introduced with a peak torque figure. Peak torque is useful for comparing ambition and poor for predicting field behavior. A joint fails its mission when it cannot repeat torque after heating, when transmission compliance destabilizes the controller, when encoder error grows under load, or when impact and cable motion degrade the module long before the motor reaches its advertised limit.
Convert the task into a torque-speed-time history
Start with representative motion, payload and contact events. Resolve joint torque and speed across time, including acceleration, gravity, coupling from other links and impact peaks. Peak values select gross mechanical capacity; RMS torque and speed-dependent losses size continuous thermal performance. Regenerative quadrants matter because the drive and shared DC bus must absorb energy rather than only deliver it.
Do not use one nominal gait. Include start-stop work, crouch, stair, recovery, manipulation and static holding. A joint with acceptable average power can still accumulate heat during a long low-speed, high-torque hold where motor current is high and mechanical output power is low.
The actuator is a coupled plant
Motor resistance and magnet flux change with temperature. Gear efficiency, backlash and friction change with load, direction, wear and lubricant condition. Housing flexibility and torsional compliance create resonances. Encoder placement determines whether the controller sees motor position, output position or both—and therefore which errors remain hidden.
Build a parameter envelope for electrical, mechanical and sensor behavior. Control tuning should be validated across that envelope, not only against a nominal identified model. A high servo bandwidth that looks precise on a rigid fixture can excite a structural mode after the actuator is mounted inside a compliant leg.
Timing is part of torque accuracy
Current sampling, PWM update, encoder acquisition, drive execution and real-time network transmission introduce delay and jitter. Whole-body control compounds the issue because synchronized joint response matters as much as one joint's local bandwidth. Measure command-to-output timing across the complete path, including EtherCAT master and distributed clocks.
Fault tests should introduce delayed frames, stale commands, encoder disagreement and drive saturation. The system needs bounded responses: hold, controlled stop, brake or torque removal according to the hazard—not an arbitrary communication timeout followed by an uncontrolled collapse.
Heat must be measured where margin is consumed
Housing temperature is not winding temperature, and drive-board temperature is not semiconductor junction temperature. Correlate accessible sensors to limiting internal points using loss models and instrumented prototypes. Test simultaneous thermal loading of neighboring joints because one link can heat another through the structure.
Acceptance should use stabilized and transient temperatures at realistic ambient, airflow and robot pose. Include battery-voltage corners because lower voltage can demand higher current for the same mechanical task and reduce voltage headroom for current control.
Reliability needs representative damage
Cycle testing should reproduce load direction, amplitude, speed, thermal state and impacts—not simply rotate the joint unloaded for a large cycle count. Track backlash, no-load current, torque constant, encoder residuals, vibration and thermal resistance as degradation indicators. These trends detect wear before catastrophic failure and become candidates for fleet health monitoring.
Release evidence
A defensible joint release includes duty-cycle traceability, torque-speed and thermal margins, servo stability across the parameter envelope, timing/jitter measurements, fault behavior, cable-flex and connector validation, and degradation trends from representative cycling. Peak torque belongs on the first page; these are the results that determine whether the robot keeps working.
Define the boundary
Name the operating state, interfaces, tolerances and degraded conditions. A test without a boundary only proves the sample survived once.
Capture the right signal
Choose measurements fast and close enough to the mechanism to explain cause, sequence and margin—not just the final alarm.
Set a decision rule
Agree before the test what passes, what triggers redesign and what uncertainty requires another experiment.
Turn the lesson into a validation plan.
Bring us the requirement, failure signature or architecture boundary. We will help frame the evidence needed to make the next decision.