Force-controlled assembly acceptance: test the task, not one force trace

The short answer: do not release a force-controlled insertion, press-fit, snap-fit or surface-following process because one trace looks smooth, the sensor datasheet is accurate or one demonstration succeeds. Build evidence through the sensor and frame chain, contact transient, steady-state and disturbance response, realistic assembly window, failure detection and safe disposition. The final gate belongs to the part: correct assembly, bounded peak load, acceptable cycle time and a defined state after failure.

Measuring a wrench, stable control and an accepted assembly are different claims

The ROS 2 Control force/torque sensor broadcaster exposes force and torque through a frame-aware, time-stamped WrenchStamped message. That proves an interface can carry data; it does not establish gravity compensation, bias, frame transforms, latency or control-loop suitability. The ATI F/T system manual explains that a tool transform translates and rotates the measurement origin to the point and axes needed by the application. A wrong TCP, sensor origin or tool orientation can therefore turn valid raw readings into a wrong contact decision.

NISTIR 8097 proposes metrics such as settle stability and disturbance handling for robots with intrinsic or extrinsic force sensing. NIST's assembly task boards then exercise system-level competencies including peg insertion, gear meshing, connector insertion and nut threading. These layers are complementary: stable contact is controller evidence; a conforming part is application evidence.

Six evidence layers from signal to product result

An executable FAT-to-SAT workflow

1. Freeze the task and conformance criteria.Record part revision and tolerance, insertion depth or functional result, tool, fixture, lubrication or surface state, cycle target and permissible load. Derive thresholds from the part, process and risk analysis—not a NIST example or sensor range.
2. Verify the measurement chain.Check calibration, units, signs, saturation status, timestamps, tool gravity and bias compensation, and the rotation and translation from sensor frame to TCP or workpiece frame. Retain raw data alongside the signal used by the controller.
3. Establish a free-space baseline.Cover representative poses, velocities, temperatures and cable-loading states. Quantify bias, noise and drift. Define when taring is permitted so that a tare taken during contact cannot erase a real load.
4. Separate contact transient from steady control.Vary approach speed and direction against a controlled artifact; record first contact, peak, overshoot, settle time and persistent oscillation. Following the intent of NISTIR 8097, add a repeatable disturbance and confirm return to the permitted band.
5. Build the real assembly window.Combine tolerance endpoints, position and angular offsets, batches, friction states and reachable fixture shifts. NIST's comparative peg-in-hole method reports completion time and probability of success while examining sensitivity to pose error and insertion force. A site test should likewise retain outcome, cycle and peak load—not success rate alone.
6. Induce failures deliberately.Cover missing and wrong parts, blocked insertion, overload, sensor saturation or loss, frame error and timeout. Confirm a bounded transition to a known state, prevention of unbounded force or search, and a diagnostic event.
7. Repeat under site conditions.Bring the FAT baseline into the final fixture, cable dress, vibration, takt time, changeover and operator workflow. Change only parameters with an identified cause and retain the FAT-to-SAT delta.
8. Release a bounded scope.Start with named part families, stations and shifts. Define which evidence layers must be repeated after changes to the tool, sensor, fixture, part material, software or critical parameters.

Minimum release evidence

• Part, tool, fixture, surface or lubrication state, tolerance and final quality criteria;
• robot, sensor, controller, software versions and force-control parameter baseline;
• calibration, tare, tool load, TCP/workpiece frames and timing-chain records;
• raw free-space, contact-transient, steady-state, disturbance and task-window traces;
• results by condition: success, cycle, peak load, part quality and intervention;
• failure state, retreat/retry logic, release scope, retest triggers and rollback parameters.

Safety boundary: ISO 9283 provides industrial-robot performance criteria and test methods, but mechanical or force-control task performance is not complete application-safety evidence. ISO 10218-2:2025 addresses integration, commissioning, operation and maintenance of industrial robot applications and cells. Contact loads, stops, safeguards, fixtures, tools and recovery actions remain part of the application-specific risk assessment and validation. A non-safety-rated wrench channel must not be promoted into a safety function by assumption.

Frequently asked questions

Does a six-axis force/torque sensor make insertion reliable?

No. Acceptance still has to cover calibration, tool gravity and bias, sensor-to-TCP or workpiece transforms, latency, control stability, and task results across real part tolerance and friction.

Is assembly success rate enough for acceptance?

No. Report the tested conditions and sample distribution together with cycle time, peak load, part quality, interventions and failure modes. A headline rate can hide damage, timeout or a very narrow operating window.

Why repeat FAT evidence during SAT?

Final fixtures, cable drag, vibration, temperature, part batches, takt time and operator workflow can change bias, friction and contact dynamics. SAT checks whether the FAT baseline still holds inside the delivered system boundary.

Can a software force threshold be treated as a safety limit?

Not by default. Only a function implemented and validated through the applicable safety architecture and standards can carry that role; an ordinary control or monitoring channel remains performance and diagnostic evidence.

Sources

These primary sources support the material facts and engineering boundaries discussed above.

1. NISTIR 8097 — Benchmarking Robot Force Control Capabilities
2. NIST — Assembly Performance Metrics and Test Methods
3. NIST — Comparative Peg-in-Hole Testing of a Force-Based Manipulation Controlled Robotic Hand
4. ROS 2 Control — Force Torque Sensor Broadcaster
5. ATI Industrial Automation — Digital F/T System Installation and Operation Manual
6. ISO 9283:1998 — Manipulating industrial robots performance criteria and test methods
7. ISO 10218-2:2025 — Industrial robot applications and robot cells