Menu

The Arm That Feels

The Arm That Feels: Force Control, Impedance Control, and Admittance

Notes · Motors and motionNovember 19, 202622 min readSenova Systems Research

Four ways to make a robot arm touch things without breaking them, direct force control, impedance, admittance, and hybrid position/force, with a real contact simulation on a bench arm.

A position-controlled arm is a blind person walking into a wall with their arms locked straight. They cannot see the wall. They cannot feel it. They keep pushing, because the command says "go there" and the only thing the controller knows is that it has not arrived yet.

That is what a robot arm does when it meets something it did not expect. The arm is told to move to a point 5 millimetres past a surface, because the surface is 5 millimetres higher than the model said, or the part is 5 millimetres thicker than the drawing said, or the person standing next to the bench moved. The position controller sees only the error, and it pushes harder to close it. The force climbs until something gives: the part, the tool, the gearbox, or the person.

Force control is the opposite skill. It is the hand that touches something in the dark and knows, from the pressure on its fingertips, whether it has found the wall, the part, or a person, and stops, or yields, or holds a steady pressure, as the task requires. The arm does not stop being precise. It stops being blind.

This article is about the four ways to give an arm that skill, and about the one decision that picks between them, a decision the hardware makes before the software ever runs.

The words, in plain English

Force control. The family of methods that regulate how hard a robot pushes, rather than only where it goes. It includes direct force control, impedance control, admittance control, and hybrid position/force control. The point of all four is the same: the arm must touch something, and the contact must be controlled.

Direct force control. The simplest member of the family: measure the contact force, compare it with the target force, and correct the difference. The target force is the setpoint, the way a target position is the setpoint in position control. It needs a force measurement, a wrist-mounted six-axis force/torque sensor, joint torque sensors, or a motor-current estimate.

Impedance control. The controller makes the arm behave like a virtual spring, damper, and mass. You choose the stiffness, the damping, and the inertia, and the arm responds to a push the way that spring would. Motion goes in, force comes out: push the arm off its path and it pushes back with a force set by the stiffness you chose. Formalised by Neville Hogan at MIT in 1985.

Admittance control. The mirror image. The controller measures the contact force and feeds it through a virtual mass-damper-stiffness model, which produces a motion command. Force goes in, motion comes out: push the arm and it yields, moving in the direction of the force by an amount set by the virtual model. The dual, the exact opposite, of impedance control.

Virtual stiffness, damping, and mass. The three numbers that define how the arm feels. Stiffness (K) is how hard it pushes back per millimetre of displacement, a spring. Damping (D) is how hard it resists motion, a shock absorber. Mass (M) is how sluggishly it accelerates, a weight. They are "virtual" because they exist in software, not in metal, but the arm behaves as if they were real.

Backdrivability. Whether an outside force can turn the joint backwards through the gearbox. A backdrivable joint yields when something pushes it. Impedance control needs it; admittance control does not.

Series elastic actuator (SEA). An actuator with a spring deliberately placed in series between the gearbox and the load. The spring deflects under load, and the deflection, measured directly, tells the controller the force, through Hooke's law. The spring is also a shock absorber and an energy store, in the same piece of metal.

Force/torque sensor. A sensor, usually mounted at the wrist, that measures three forces and three torques. It is the admittance controller's eyes: without it, there is no measured force to feed the virtual model.

Contact stiffness. How stiff the thing being touched is. A steel table is stiff; a foam pad is soft. The contact force for a given position error is the error times the contact stiffness, which is why a 1-millimetre error against steel is a different event from a 1-millimetre error against foam.

Hybrid position/force control. The method that splits the task by direction: control position along some axes and force along others. Polishing a surface means moving along the surface in position control while holding a steady force into the surface in force control.

Selection matrix. The table in a hybrid controller that says which direction gets position control and which gets force control. It is the formal version of the decision above.

Natural and artificial constraints. Mason's framing, which Raibert and Craig built on. Natural constraints are what physics dictates, a surface prevents motion into it. Artificial constraints are what the task dictates, the polish must press at 5 newtons. The hybrid controller assigns each direction to the constraint that governs it.

Quasi-static vs transient contact. The two kinds of contact in the safety standard. Quasi-static contact is clamping or trapping, a body part pressed between the robot and a fixed object, and held there. Transient contact is a brief impact, the body part can recoil. The limits are different, because the physics is different.

Passivity. The property that a controller never adds energy to the interaction. A passive controller can only absorb energy or return what it stored, it cannot push harder than the push it received. Passivity is what makes a controller safe to couple to an unknown environment, because an unknown environment cannot be made unstable by a controller that never adds energy.

Damping ratio. How quickly an oscillation dies away, as a fraction of critical damping. A ratio of 1.0 is critically damped, it returns to rest without overshooting. A ratio of 0.7 to 1.0 is the usual target: the disturbance decays without bouncing.

The problem: position control in a stiff environment

The default control mode on most industrial robots is stiff position control: track the commanded trajectory regardless of external forces. It is the right mode for free-space motion, where contact is not expected. During contact, it is the wrong mode, and the arithmetic shows why.

A stiff position controller is a very stiff spring. Its stiffness is the position gain: a gain of 50,000 newtons per metre means a 1-millimetre position error produces 50 newtons of force, and a 10-millimetre error produces 500 newtons. The tolerances in a real assembly are measured in fractions of a millimetre, peg-in-hole insertion is typically ±0.5 millimetres, a USB-A connector ±0.3 millimetres, a SIM card slot ±0.1 millimetres, and the parts, the fixtures, and the arm itself are never exactly where the model says they are.

A 1-millimetre error during a 0.5-millimetre-clearance insertion produces contact forces that can reach hundreds of newtons.

That is the failure mode. The arm does not know it has touched anything. It knows only that it has not arrived, so it pushes harder. The force climbs until something gives. The demo below runs the numbers on our bench arm, and the stiff controller's peak contact force is 1,159 newtons, more than eight times the 140-newton limit the safety standard sets for a hand.

The fix is not a lower gain. A lower gain makes the arm sloppy in free space and slow to settle. The fix is a controller that knows about force, and there are four of them.

The four families, and the one question that picks between them

Every force-control method answers the same question differently: which is the input, and which is the output, motion or force?

MethodInputOutputBehaves likeNeeds
Direct force controlMeasured force errorMotion or torque correctionA thermostat for forceA force measurement, and a loop fast enough to track it
Impedance controlMotion errorForce/torqueA spring (and damper)Backdrivable, torque-controlled actuators
Admittance controlMeasured forceMotion referenceA damper (it yields)Stiff actuators, a position loop, and a force/torque sensor
Hybrid position/forceDirection, split by taskPosition in some directions, force in othersA hand on a surfaceA selection matrix, and a force measurement

Direct force control: the thermostat

Direct force control makes the target force itself the feedback setpoint. The controller continuously compares the target force with the actual force and generates corrective commands to reduce the difference, the way a thermostat compares the target temperature with the measured temperature and switches the heater.

It is the most literal member of the family, and the most demanding. It needs a real force measurement, and the loop has to be fast enough to track the contact: sensor noise, contact stiffness, actuator bandwidth, and delays all feed directly into closed-loop stability. It is the right tool when the task has a force target and nothing else matters, a press that must hold 50 newtons, a grip that must not crush the part.

It is the wrong tool when the task also needs the arm to go somewhere precise, because a pure force loop cannot stably position even a mass on its own, Hogan showed that in 1985. A force loop with nothing to hold it in place drifts.

Impedance control: motion in, force out

Impedance control does not command a force. It commands a relationship between motion and force, and the relationship is a virtual spring, damper, and mass:

F = M·(a_d − a) + D·(v_d − v) + K·(x_d − x)

Push the arm off its desired path by (x_d − x), and it pushes back with a force set by the stiffness K. Move it fast, and the damping D resists. Accelerate it, and the virtual mass M resists. The controller never says "apply 5 newtons." It says "behave like a spring of 500 newtons per metre", and the force that results depends on how far the environment pushes the arm off its path.

Hogan's 1985 paper, "Impedance Control: An Approach to Manipulation," Parts I to III, in the ASME Journal of Dynamic Systems, Measurement, and Control, is the foundation. His argument is the one the whole field rests on: a manipulator in contact cannot be controlled by commanding position or force alone; it must also be given an impedance, the force it produces in response to motion the environment imposes. And because the environment contains inertias and kinematic constraints, the environment is an admittance, so the manipulator should be an impedance.

The two forms are not interchangeable: along each degree of freedom, a system either sets force from motion (an impedance) or sets motion from force (an admittance), and the choice is forced by what the environment is.

Impedance control is the standard approach for contact tasks on torque-controlled robots: assembly, polishing, human-robot collaboration, walking robots touching the ground. It absorbs impacts well, because the actuator is backdrivable and the torque loop is fast. Its weakness is the one the equation shows: it does not guarantee a specific contact force. The force depends on the displacement, and the displacement depends on the environment, which the controller does not know. If the task needs exactly 5 newtons, impedance control alone cannot promise it.

Admittance control: force in, motion out

Admittance control is the exact dual. The outer loop reads a force/torque sensor, feeds the measured force through a virtual mass-damper-stiffness model:

M_a⋅x¨_r+D_a⋅x˙_r+K_a⋅(x_r−x_0)=F_ext\text{M\_a·ẍ\_r} + \text{D\_a·ẋ\_r} + \text{K\_a·(x\_r} - \text{x\_0)} = \text{F\_ext}

and the model produces a motion reference, a position or velocity, which an inner position or velocity loop then tracks. Push the arm with 5 newtons, and the virtual model moves the reference by 5 ÷ K_a, 10 millimetres, if K_a is 500 newtons per metre. The arm yields.

Admittance control is the right tool when the robot already has a strong inner position or velocity loop but cannot expose direct low-level torque control, which is most industrial arms and most cobots. It preserves payload density, because it keeps the stiff, high-ratio gearboxes. Its weakness is the impact transient, and the demo shows it plainly: the stiff inner loop races to the target before any force is measured, so the first hit is as hard as a position controller's.

The admittance loop can only yield after the force has been measured, and by then, the contact has already happened.

Its stability is also the most fragile of the four. Keemink, van der Kooij, and Stienen's 2018 review in the International Journal of Robotics Research, "Admittance control for physical human–robot interaction", analyses seven factors that decide whether an admittance loop stays stable: feed-forward control, force-signal filtering, post-sensor inertia compensation, virtual damping, phase lead on the motion reference, motion-loop bandwidth, and internal robot flexibility. The short version: the stability margin shrinks when sensor delay, filtering, and a stiff environment add phase lag.

A low virtual mass makes the arm highly responsive to touch but prone to vibration; a high virtual mass makes it heavy but stable.

Hybrid position/force: split by direction

Hybrid position/force control, from Raibert and Craig's 1981 paper, splits the task by direction instead of by method. It assigns force objectives to some task-space directions and position objectives to others, through a selection matrix, in a task-related coordinate frame. The classic example is surface polishing: control X and Y in position mode (follow the surface trajectory) while controlling Z in force mode (maintain 5 newtons of contact force against the surface). Another is writing with a pencil: position along the page, force into the page.

The framework comes from Mason's theory of natural and artificial constraints: the surface prevents motion into it (a natural constraint, dictated by physics), and the task demands a pressure (an artificial constraint, dictated by the goal). The hybrid controller assigns each direction to the constraint that governs it, so the control action never conflicts with the geometry. It is the most task-shaped of the four methods, and the most honest about what a contact task actually is: some directions are about where the tool goes, and some are about how hard it presses.

The hardware decides before the software does

Here is the decision the article exists to make, and it is made by the actuator, not by the programmer.

Impedance control needs backdrivable actuators: quasi-direct drive, low gear ratios, accurate current sensing. The controller commands torque, and the torque has to reach the joint, which means the gearbox must not stand in the way. A 100:1 harmonic drive, which reflects the motor's inertia by a factor of 10,000 and is effectively not backdrivable, cannot do impedance control well. The gearbox article's numbers are the reason: the harmonic's backdrive torque is 5–25 newton-metres against a quasi-direct drive's 0.3–1.5.

Admittance control needs the opposite: stiff actuators, high gear ratios, and direct force/torque sensing. It keeps the harmonic drive and the payload density, and it closes the loop around a wrist-mounted sensor instead of around the motor current. The sensor is not optional, without a measured force, there is nothing to feed the virtual model.

The middle path is the series elastic actuator. Pratt and Williamson, at MIT in 1995, put a spring deliberately in series between the gearbox output and the load. The spring does four jobs in one piece of metal: it absorbs shock before it reaches the gear train; it stores and releases energy during cyclic motion, the way a tendon does; it acts as a low-pass mechanical filter, smoothing torque ripple from the motor and gearbox; and, the point for this article, it is a force sensor.

Measure how much the spring deflects, and Hooke's law gives the force: force equals stiffness times deflection. No external force sensor is needed, and the force reading is direct and low-noise, because the spring's deflection is the force. The price is bandwidth: the spring trades achievable control bandwidth for stable, low-noise force control.

A measured torsional series-elastic element, designed by finite-element analysis, reached a stiffness of 2,155.4 newton-metres per radian, and one degree of deflection on that spring is 37.6 newton-metres of torque.

The rule, then: if the actuator is backdrivable and torque-controlled, impedance. If the actuator is stiff and position-controlled with a force/torque sensor, admittance. If the task splits cleanly by direction, hybrid. If the force has a single target and nothing else matters, direct force control. The two architectures are not interchangeable, sensor placement, inner-loop access, task geometry, and contact dynamics decide which fits, and choosing the wrong one is not a tuning problem. It is a hardware problem, and no amount of gain tuning fixes a hardware problem.

The safety numbers

Force control is also a safety device, and the safety standard gives it numbers. ISO/TS 15066, the technical specification for collaborative robots published in 2016, sets maximum permissible contact forces for 29 body locations in 12 body regions, based on pain-onset research, not injury thresholds, which is why the numbers are deliberately conservative for routine contact. The quasi-static limits, clamping or trapping, where the body part cannot move away, run from 65 newtons for the face to 220 newtons for the thighs and knees.

The hand and fingers, the region a bench arm is most likely to touch, are 140 newtons. The chest is 140 newtons. The skull and forehead are 130 newtons.

Transient contacts, a brief impact the body part can recoil from, are allowed up to twice the quasi-static limit, because the energy is absorbed over a brief period.

The standard also gives a speed formula, which turns a force limit into a speed limit: the allowed end-effector speed depends on the force limit, the body region's spring constant and effective mass, and the robot's effective mass. The demo below uses the force limits directly, because they are the number the controller has to respect: a controller whose steady contact force is 83 newtons is inside the 140-newton hand limit, but a controller whose peak is 1,159 newtons is not, and the peak is the number that matters in the first 30 milliseconds.

Tuning: the three numbers, and the one rule

The virtual stiffness, damping, and mass are the tuning knobs, and the guides agree on the ranges. Translational stiffness on collaborative arms runs about 100 to 5,000 newtons per metre. Low stiffness, 50–200 newtons per metre, makes the robot very compliant, useful for contact exploration, where the arm has to find the surface without hurting it. High stiffness, 2,000–5,000 newtons per metre, makes it nearly rigid while still protecting against force spikes. Typical manipulation work sits at 200–1,000 newtons per metre for translation and 10–50 newton-metres per radian for rotation.

Damping is usually chosen relative to stiffness to obtain near-critical damping, so a disturbance decays without overshoot: D = 2·√(K·m), with a damping ratio of about 0.7 to 1.0. Too little damping and the arm bounces; too much and it is sluggish. And one rule from the human-robot interaction literature: leave the integral term out. During physical interaction, an integral term accumulates error while the arm is pressed against a surface, integral wind-up, and then exerts excessive force and overshoots when contact breaks.

That is dangerous, and it is the reason the impedance loop in the demo below is a PD law, not a PID law.

Example Labs is a made-up company we use for worked examples.

the Example Labs bench arm meets a wall

The arm is the same one from the motor and gearbox articles: two links, each 0.425 metres long, a 5-kilogram payload at 0.85 metres reach, the shoulder needing 126.8 newton-metres with the 1.5× safety margin. The demo is a one-degree-of-freedom contact simulation, run in Python before writing: a 2-kilogram effective mass at the end of the arm, a rigid wall at 100 millimetres, the arm commanded to 105 millimetres, 5 millimetres into the wall, the position error a real assembly would produce. Three controllers, same wall, same command.

Step 1: the arithmetic of a stiff position controller

Before the simulation, the arithmetic. A position controller's force is its gain times its error:

  • 1 millimetre of error at 50,000 N/m = 50 N
  • 10 millimetres of error at 50,000 N/m = 500 N
  • 1 millimetre of error at 5,000 N/m = 5 N

The ISO/TS 15066 hand limit is 140 newtons. At a gain of 50,000 N/m, the arm reaches it with 2.8 millimetres of error. At 20,000 N/m, it takes 7 millimetres. The tolerance on a peg-in-hole insertion is ±0.5 millimetres, and the parts are never exactly where the model says they are. The margin between "in tolerance" and "over the safety limit" is a few millimetres, and the controller cannot see it coming, because it cannot feel.

Step 2: the simulation: three controllers, one wall

ControllerApproach speed at contactPeak contact forceSteady contact force
Stiff position control (Kp 20,000 N/m)3,583 mm/s1,159 N at 28.5 ms83.3 N
Impedance control (K 500 N/m, D 50 N·s/m)169 mm/s71.3 N at 219 ms7.0 N
Admittance control (M 2 kg, D 50, K 500, inner loop Kp 20,000)3,583 mm/s1,151 N at 28.4 ms0.0 N

The position controller arrives at 3.6 metres per second and hits the wall with 1,159 newtons, more than eight times the 140-newton hand limit, and then keeps pushing at 83 newtons, because the command still says 105 millimetres and the wall is at 100. The impedance controller arrives at 169 millimetres per second, because the soft spring accelerates the mass gently, and its peak is 71 newtons, inside the limit, settling at 7 newtons, which is the spring's force at the 5-millimetre penetration (500 N/m × 0.005 m = 2.5 N, plus the wall's own compliance in the simulation).

The admittance controller is the interesting one. Its peak is 1,151 newtons, as hard as the position controller's, because its stiff inner loop races to the target before any force is measured, and the outer loop can only yield after the force has been measured. The impact transient is the gap in admittance control, and no tuning of the virtual model closes it: the force has to exist before the loop can react to it.

But look at the steady state: 0.0 newtons. The virtual model retreats the reference until contact breaks, and the arm ends up 100.00 millimetres from the wall, touching it, pushing nothing. The admittance controller cannot prevent the first hit.

It can end the contact.

Step 3: the hybrid task: polishing a surface

The classic hybrid task, from the literature: polish a surface. Control X and Y in position mode (follow the surface trajectory) and control Z in force mode (maintain 5 newtons of contact force against the surface).

The force loop is a direct force controller on the Z axis: the force error is the target minus the measured force, and the motion correction is the error divided by the Z stiffness. With a Z stiffness of 500 N/m, a 0.5-newton force error produces a 1-millimetre correction, the arm presses a millimetre deeper when the force is half a newton light, and lifts a millimetre when it is half a newton heavy.

The surface does not have to be flat. A height difference of a few millimetres would make a position-controlled arm either miss the board or press too hard; the hybrid controller keeps the pressure at 5 newtons and lets the Z position follow the surface. That is the whole point of splitting by direction: the position axes do the travelling, and the force axis does the touching.

Step 4

Take the measured torsional series-elastic element from the literature: a spring stiffness of 2,155.4 newton-metres per radian, designed by finite-element analysis. One degree of deflection on that spring is 37.6 newton-metres of torque, measured directly, from the deflection, with no force sensor anywhere in the joint.

Or the linear version, for the polishing task: a spring of 5,000 newtons per metre deflects 1 millimetre under the 5-newton polishing force. The controller reads the deflection, multiplies by the stiffness, and knows the force. The spring also absorbs the impact that a rigid gearbox would transmit straight to the motor, which is why series elastic actuators were invented for robots that interact with people, where rigid force spikes can cause injury.

Step 5

The ISO/TS 15066 hand limit is 140 newtons, quasi-static. Against it:

  • Position control: peak 1,159 N, eight times the limit. It fails the safety case, and the failure happens in the first 30 milliseconds, before any software limit can react.
  • Impedance control: peak 71 N, steady 7 N, inside the limit on both counts. It passes, because the soft spring limits the force and the slow approach limits the impact.
  • Admittance control: peak 1,151 N, it fails the transient, because the stiff inner loop hits the wall before the outer loop knows the force exists. Steady 0 N, it passes the sustained contact, because it yields until the force is gone.

The lesson is not that admittance control is unsafe. It is that admittance control needs a second layer for the transient, a slower approach, a force limit on the inner loop, or an impact-absorbing element in the joint, and the safety case has to be written for the peak, not the steady state. A controller that ends at zero force can still hurt someone in the first 30 milliseconds.

The decision test

Four questions, in order. If you cannot answer the first, the rest do not matter.

  1. What is the contact task, in one sentence? "Hold 5 newtons while moving along the surface" is a hybrid task. "Yield when pushed" is admittance. "Behave like a spring of 500 newtons per metre" is impedance. "Hold 50 newtons, nothing else matters" is direct force control. If the task cannot be said in one sentence, the controller cannot be chosen.
  2. Can the actuator command torque, or only position? If the joint is backdrivable and torque-controlled, quasi-direct drive, low ratio, current sensing, impedance is available. If the joint is stiff and position-controlled with a force/torque sensor, a harmonic drive and a wrist sensor, admittance is the architecture. This is a hardware question, and the answer is in the gearbox article, not in the code.
  3. What is the force limit, and which number enforces it? The ISO/TS 15066 limit for the body region the arm can touch, 140 newtons for a hand, 65 for a face, and the controller that enforces it. Impedance enforces it with stiffness. Admittance enforces it with the virtual model, after the transient. Direct force control enforces it with the setpoint. If no number enforces it, the limit is a hope, not a design.
  4. What happens in the first 30 milliseconds? The impact transient is the gap in admittance control and the failure in position control. Run the contact simulation, the one in this article, on your arm, with your wall, and read the peak, not the steady state. If the peak is over the limit, the approach speed, the stiffness, or the joint needs to change.

Objections

"But impedance control cannot hold an exact force." Correct, and the article says so. The force depends on the displacement, and the displacement depends on the environment. If the task needs exactly 5 newtons, use direct force control or hybrid control for that axis, and use impedance for the rest. The methods are not rivals. They are tools, and the hybrid controller exists precisely because one tool is rarely enough.

"But admittance control is just impedance control with the loop turned inside out." They are duals, and they implement the same virtual dynamics, but the causality is the whole difference. Impedance needs torque control and backdrivability; admittance needs a position loop and a force sensor. The hardware that runs one well runs the other badly, and the impact transient in the demo is the proof: the same virtual model, the same stiffness, and one peaks at 71 newtons while the other peaks at 1,151.

"But our arm is position-controlled, so we cannot do force control." Your arm is position-controlled, so you can do admittance control, it is built for exactly that architecture, and hybrid control, and direct force control on one axis. What you cannot do is impedance control, because that needs torque control the arm does not expose. The question is not whether you can do force control. It is which family your hardware supports.

"But the safety standard's limits are just guidance." ISO/TS 15066 is a technical specification, not a full standard, and it is being folded into ISO 10218:2025. But the numbers come from pain-onset research, and they are the numbers a risk assessment will be judged against. A controller whose peak is eight times the hand limit does not have a safety case, whatever the document's status.

Sources

  1. Hogan, N., "Impedance Control: An Approach to Manipulation: Part I, Theory"
    What we used

    (ASME Journal of Dynamic Systems, Measurement, and Control, vol. 107, no. 1, pp. 1–7, 1985; Parts I–III, pp. 1–24), "control of a vector quantity such as position or force is inadequate and control of the manipulator impedance is also necessary"; the environment contains inertias and kinematic constraints, so it is an admittance, and the manipulator should be an impedance; the distinction between admittance and impedance is fundamental; along each degree of freedom a system either sets force from motion (an impedance) or sets motion from force (an admittance), and the two forms are not interchangeable; pure force control cannot stably position even a mass

  2. Keemink, A. Q. L., van der Kooij, H., and Stienen, A. H. A., "Admittance control for physical human–robot interaction"
    What we used

    (International Journal of Robotics Research, vol. 37, no. 11, pp. 1421–1444, 2018), admittance control as the opposite, or dual, of impedance control; the framework analyses feed-forward control, force-signal filtering, post-sensor inertia compensation, virtual damping, phase lead on the motion reference, internal robot flexibility (which relates to series elastic control), motion-loop bandwidth, and virtual damping on the stability, passivity, and performance of minimal-inertia rendering admittance control; seven design guidelines for high-performance admittance devices

  3. Raibert, M. H., and Craig, J. J., "Hybrid Position/Force Control of Manipulators"
    What we used

    (ASME Journal of Dynamic Systems, Measurement, and Control, vol. 102, June 1981), combines force and torque information with positional data to satisfy simultaneous position and force trajectory constraints, using feedback from a force-sensing wrist and joint position sensors; based on Mason's theory of natural and artificial constraints

  4. Pratt, G., and Williamson, M., "Series Elastic Actuators"
    What we used

    (IEEE/RSJ International Conference on Intelligent Robots and Systems, 1995), a spring placed deliberately in series between the gearbox output and the load; deflection gives force through Hooke's law; low output impedance, shock tolerance, stable low-noise force control; developed for robots that interact with people

  5. ISO/TS 15066:2016, "Robots and robotic devices, Collaborative robots"
    What we used

    Maximum permissible force and pressure for 29 body locations in 12 body regions, based on pain-onset research; quasi-static limits: face 65 N, skull/forehead 130 N, chest 140 N, hand/fingers 140 N, thigh/knee 220 N; transient contacts allowed up to twice the quasi-static limit; speed formula v ≤ F_max/√k × √(1/m_R + 1/m_H); being folded into ISO 10218:2025

  6. Robotics Center, "Why Contact Forces Are the Key to Dexterous Robot Manipulation"
    What we used

    Stiff position control drives high forces to maintain the commanded position; a 1 mm position error during a 0.5 mm-clearance insertion produces contact forces that can reach hundreds of newtons; peg-in-hole ±0.5 mm, USB-A ±0.3 mm, SIM card ±0.1 mm; low stiffness 50–200 N/m for contact exploration, high stiffness 2,000–5,000 N/m nearly rigid; typical manipulation K_translation 200–1,000 N/m, K_rotation 10–50 N·m/rad, D = 2·√(K·m); hybrid example, surface polishing, X/Y in position, Z in force at 5 N

  7. Webiano, "Impedance control, Robotics"
    What we used

    F = K(x_d − x) + D(v_d − v) + M(a_d − a); translational stiffness about 100–5,000 N/m on collaborative arms; damping ratio about 0.7–1.0; impedance control does not guarantee a specific contact force; impedance measures motion and commands force, admittance measures force and commands motion

  8. Humanoid Joint, "Admittance control vs impedance control"
    What we used

    And "Admittance vs impedance control guide", impedance for dynamic legged robots or high-impact joints only when the actuator stack is backdrivable, torque-controllable, and validated at the target loop rate; admittance for stiff, highly geared arms or cobots when direct force/torque sensing can close the force-to-motion loop; admittance stability margin shrinks when sensor delay, filtering, and a stiff environment add phase lag; low virtual mass highly responsive but prone to vibration, high virtual mass heavy but stable

  9. OSREC, "Impedance and Admittance Control"
    What we used

    Impedance maps motion error into force/torque; admittance maps measured force/torque into motion; torque-controlled, backdrivable arms suit direct impedance control; position-controlled, high-ratio robots suit admittance when an external force/torque sensor is available

  10. Embodied Terms, "What is Force Control?"
    What we used

    Pure position control responds to any deviation by pushing harder, so if it meets a surface higher than expected it can crush the workpiece; direct force control makes the target force itself the feedback setpoint; indirect force control (impedance and admittance) sets a stiffness, damping, and inertia; the whiteboard-wiping example, a height difference of a few millimetres makes a position-controlled arm miss the board or press too hard

  11. Article #3, "PID Tuning, Simulation First" (Senova, 2026-11-19)
    What we used

    The inner loop the force loop wraps around; the simulation-first discipline

  12. Article #10, "Sense, Decide, Move" (Senova, 2026-11-19)
    What we used

    The arm, and the human-approval gate

  13. Article #21, "The Gearbox Is the Joint" (Senova, 2026-11-19)
    What we used

    Harmonic backdrive torque 5–25 N·m vs quasi-direct drive 0.3–1.5 N·m; reflected inertia J = J_motor × N²; the shoulder 126.8 N·m, elbow 40.7 N·m, wrist 11.0 N·m

  14. Article #20, "Choosing the Motor" (Senova, 2026-11-19)
    What we used

    The motor torques the gearbox multiplies; quasi-direct drive architecture

More in motion and sensing

The Axis That Doesn't BacklashA linear axis misses position, whips at speed, or drops its load on power loss, and the fault is usually the architecture, not the motor. Ball screw vs lead screw vs belt vs rack vs linear motor, the five numbers that decide, and a worked Z-axis sizing.16 minThe Drive StageA BLDC driver is five sections, six switches, and one rule that matters more than the rest: the two devices in a half-bridge must never conduct together. This is how the drive stage works, how it dies, and the arithmetic that keeps it alive, with a worked 48 V design.30 minThe Encoder Is the LoopAn encoder is the sensor that tells the controller where the shaft is, that is the whole job. Incremental counts pulses and forgets everything on power loss; absolute reads a unique code for every angle and knows where it is the instant power returns. This is what resolution buys, why it is not accuracy, why velocity comes from differentiating position and gets noisy at low speed, and what Example Labs' bench arm needs on each of its four joints.24 minThe Gearbox Is the JointPlanetary, harmonic, and cycloidal gearboxes compared on the four numbers that decide a robot joint, backlash, stiffness, efficiency, and cost, with a worked demo on a real arm.18 minThe Motor That Doesn't HumYour motor hums because the drive is switching in blocks, and the torque dips at every 60° boundary. Field-oriented control holds the current exactly 90° ahead of the rotor flux, so the ripple falls from about 14% to under 1%, and the harmonics stop becoming heat. Here is the arithmetic, and the one place trapezoidal still wins.27 min