Impedance and Admittance Control, Without the Math

A robot arm that always tries to reach a commanded pose regardless of resistance is called stiff. A robot arm that yields under external force in a controlled way is called compliant. The difference is one of the most important design decisions in manipulation, and it has recently become one of the most important design decisions in policy learning as well.

The vocabulary around compliance, impedance control, admittance control, hybrid force-position, is technical enough to keep most product conversations out of it. It should not be. The concepts are simple, the implications are large, and any conversation about foundation-model deployment sooner or later has to touch them.

An impedance controller makes the robot behave like a mass on a spring and damper. You tell it a target pose. It measures how far it is from that pose. It applies force proportional to the offset, minus some damping proportional to the velocity. If nothing is in the way, it moves smoothly to the target. If something pushes back, it yields, because the spring metaphor is real: the further the deviation, the more force, but never infinite force. The stiffness of the spring is a tunable parameter, and low stiffness means a compliant robot.

An admittance controller flips the input and output. You measure the external force on the robot, decide how the robot should move in response, and command that motion. It is what you use when the robot's own actuators are stiff and you want to add apparent compliance in software. Admittance and impedance often produce similar behavior. They diverge in edge cases, especially near contact transitions, and choosing between them is a nontrivial engineering decision.

Every task that involves inserting one thing into another, wiping a surface, holding a soft object, or coordinating two hands is a compliance problem. Stiff position control succeeds only if the world matches the model exactly. It fails the moment the peg is a millimeter off, or the surface is uneven, or the object deforms under contact. The failure modes are ugly: high forces, damaged parts, jammed grippers.

A compliant robot behaves differently. It yields to unexpected contact rather than fighting it. It follows a surface rather than gouging it. It holds a fragile object with a defined force rather than a defined position. These behaviors are the physical prerequisites for most of the manipulation the field wants to automate, and they are unavailable to any control stack that treats the world as rigid.

A learned policy sits on top of a lower-level controller. If the lower-level controller is stiff position control, the policy inherits stiffness. It can command trajectories, but at the moment of contact the underlying stiffness dominates. If the lower-level controller is compliant, the policy has a chance to specify richer behaviors, target contact forces, allowable position drift, direction-specific stiffness, that the underlying controller can execute.

The catch is that learned policies can only specify behaviors they have seen. A policy trained on demonstrations that came from stiff position control has learned to plan in position, and it will fight compliance rather than exploit it. A policy trained on demonstrations from compliant systems, including compliant humans, has learned to plan in force and position simultaneously, and it can specify the compliance parameters as part of the action.

The data problem is exactly the one you would expect. Most robot training data was collected on stiff position-controlled setups, so most trained policies are stiffness-blind. The datasets that would teach compliance are the ones where the demonstrator was themselves compliant, meaning either a compliant robot in a well-designed teleoperation loop or, more directly, a human hand.

Human hands are exquisite compliance instruments. The whole hand is a distributed spring-and-damper system, tuned unconsciously to the task, and every skilled manipulation trajectory is implicitly a compliance demonstration. Capturing that trajectory with the right sensors preserves the compliance information. Capturing it without force preserves only the outer shape.

The next two pieces trace this thread into the specific reason compliance policies remain undertrained and how a human-first capture pipeline changes it.