Gravity compensation

Balance a configuration-dependent load#

The gravitational torque at each joint changes with the robot's configuration, link masses, centers of mass, and payload. A model can predict that torque and add it to the command. In an ideal static case, the compensation supplies the effort needed to hold the mechanism against gravity without requiring a position error.

MIT's manipulator-control notes show why a proportional-derivative controller alone needs a nonzero error to produce a steady torque under gravity, then introduce model-based gravity compensation. The compensation is feedforward: measured position selects a modeled load rather than proving that the load was canceled.

Holding and hand-guiding are different goals#

A position controller can combine gravity compensation with feedback so that the robot tracks a pose without sagging. A hand-guided robot can instead use compensation with low commanded stiffness so that an operator does not have to lift the robot's full weight.

Active gravity compensation is not the same as mechanical counterbalancing with springs or counterweights. It also does not remove inertia, friction, gearbox drag, or controller stiffness. A powered joint can feel easy to lift while still resisting fast motion. This is why backdrivability should be assessed separately.

Model errors leave residual forces#

An incorrect payload, center of mass, joint zero, or base orientation changes the true gravity torque. The residual can cause drift, pose error, or an unexpected force against a person or object. Saturation and communication delay add further error on large mechanisms.

Gravity compensation is a narrower model term than computed-torque control. The latter can also account for inertia, velocity-dependent effects, and desired acceleration. Modern Robotics notes that gravity compensation plus feedback is cheaper to evaluate than a full dynamic controller, but it describes fewer effects.