Compliant Control & Force-Limiting Architecture
Duration: 55 min · Level: Intermediate · Module: 8. Safety & Human-Robot Interaction · Focus: compliance, force-control, impedance, safety
A force limit on a datasheet is a promise; making the robot keep it in the messy real world is an engineering problem. Lesson 8.1 gave you the numbers G1 must not exceed. This lesson builds the hardware-and-software stack that actually senses contact and clamps force in real time — the machinery behind ISO/TS 15066's power-and-force-limiting mode. The recurring theme: you have several ways to feel a collision, each with a cost, and the right design layers them so that no single sensor failure removes your protection.
Sensing contact: where the force signal comes from
You cannot limit a force you cannot measure, so the first decision is how G1 senses contact. There are three practical sources, and they form a hierarchy of cost and accuracy.
Joint torque sensing is the cheapest and most pervasive option. Using strain gauges at each joint — or simply reading motor current, since current is proportional to torque — you can estimate the external force acting on the limb from the proprioceptive signals the robot already produces. The honest limitation: with proprioceptive sensing alone, accuracy lands around ±2–5 N. That is fine for catching a torso bump against the 65 N thorax limit, but marginal if you need to handle a delicate object at a few newtons.
Wrist force/torque (F/T) sensors sit at the other end. A 6-DOF F/T sensor mounted at each wrist measures all three forces and three torques at the end-effector directly, giving accurate force readings exactly where most manipulation contact happens. Off-the-shelf units like the ATI Mini45 (or a custom equivalent) are the standard choice. The cost is money, mass at the worst possible place (the far end of the arm), and another cable to route — but for manipulation safety, the accuracy is worth it.
The recommendation: use both, deliberately. Joint torque sensing gives you whole-arm contact awareness cheaply — it notices a collision anywhere along the link, not just at the tool. Wrist F/T gives you high-fidelity end-effector force for the delicate tasks. They are complementary, not redundant: the wrist sensor is blind to a contact at the elbow, and the joint estimate is too coarse for fine manipulation. Layer them.
Reacting to contact: impedance control instead of position control
Sensing contact is only half the loop; the robot also has to respond gently. A position-controlled arm fights any deviation from its commanded path with full authority — exactly the wrong instinct when the deviation is a human.
The alternative is impedance control: instead of commanding a position and rejecting all disturbance, you regulate the relationship between displacement and force — the endpoint's effective stiffness and damping. Tell the controller "behave like a soft spring," and when it meets resistance it yields proportionally rather than pushing through. Low stiffness ("soft" impedance) makes contact inherently gentle, which is why it underpins cobots like the Franka Emika Panda. The trade-off is precision: a soft arm tracks position less tightly, so you tune stiffness per task — stiffer for precise insertion in free space, softer the moment contact with a person is possible.
Impedance control pairs naturally with the backdrivable, low-gear-ratio actuators from Module 2. A joint that is easy to backdrive can physically yield to force; impedance control is how you command that yielding intelligently rather than leaving it to mechanics alone.
Catching collisions you didn't see coming
Even with torque sensing and soft impedance, you want a fast, sensor-light tripwire for genuinely unexpected impacts. The standard tool is the momentum observer: it continuously predicts what the joint velocities should be given the commanded torques and the robot's dynamics, and compares that to the actual velocities. A discrepancy means an external force is acting that the model didn't account for — a collision. Crucially, this works without any external sensor, using only the robot's own state, and it detects contact in under 10 ms — fast enough to halt before a transient impact climbs toward the injury threshold.
The momentum observer is your independent safety net. The wrist F/T sensor might be unplugged, the impedance gains might be mistuned, but a model-based observer running on the robot's own signals still catches the collision. Build it as a separate watchdog, not as a feature buried inside the main controller.
Don't wait for contact — predict it
The fastest collision response is the one you never need. Predictive safety moves the line of defense earlier: the trajectory planner checks every future configuration of the robot against a model of the workspace and refuses — or stops — motion before the robot would enter a region known to be occupied. Where the momentum observer is reactive (something already touched me), predictive checking is proactive (that path would touch something, so don't take it). This layers cleanly on top of everything above: predict to avoid contact, observe to catch the contact you couldn't predict, impedance-control to soften the contact that lands, and torque-limit to cap its force.
The four layers compose into one defense-in-depth loop — each catches what the one before it missed:
The risk assessment ties it together
None of this is certifiable without the ISO 10218-2 risk assessment — the document that, for every hazard scenario, estimates probability and severity and records the mitigation applied until residual risk is acceptable. This is not paperwork bolted on at the end; it is the design discipline that tells you which of the above layers each hazard actually needs. A pinch hazard at the elbow might be addressed by geometry and the momentum observer; a crushing hazard during medication handoff might demand the wrist F/T sensor plus soft impedance. The assessment is what a CE/UL auditor reads, and it is how you prove your layered architecture is sufficient rather than merely impressive.
Putting it into practice
Specify G1's force-limiting stack for the medication-handoff motion and sketch the momentum-observer logic.
- For the handoff, choose the sensing sources: confirm joint torque sensing for whole-arm coverage and wrist F/T for the gripper, and write the accuracy you expect from each (≈±2–5 N proprioceptive; tighter at the wrist).
- Pick an impedance setting: high stiffness while the arm moves through empty space, dropping to low stiffness once the gripper is within reach of the patient. State the trigger that switches modes.
- Write the momentum-observer rule in one line: if (predicted joint velocity − actual joint velocity) exceeds a threshold, trigger a safety stop — and require it to fire in under 10 ms.
- Add one predictive check: before executing, verify the planned trajectory never places any link inside the patient's modeled occupied region.
- Open a one-row risk-assessment entry for "unexpected contact during handoff": list the hazard, its severity against the 65 N limit, and which of the four layers above mitigate it.
Key takeaways
- Sense contact from joint torque (cheap, whole-arm, ±2–5 N) and wrist 6-DOF F/T sensors (accurate end-effector force, e.g. ATI Mini45) — they are complementary, so use both.
- Impedance control regulates endpoint stiffness/damping instead of position; soft (low) stiffness makes contact gentle, as in the Franka Emika Panda, and pairs with backdrivable actuators.
- A momentum observer detects collisions in <10 ms using only the robot's own signals — build it as an independent watchdog with no external sensor dependency.
- Predictive trajectory checking stops motion before entering occupied regions, layering avoidance ahead of reaction.
- The ISO 10218-2 risk assessment drives which mitigation each hazard needs and is the document auditors review — treat it as design discipline, not afterthought.
← Previous: 8.1 ISO 10218 & ISO/TS 15066: The Collaborative Robot Safety Standards · Next: 8.3 Behavior Trees & Task-Level Safety →
Part of Module 8: Safety & Human-Robot Interaction.