computed-torque control

Computed-torque control is a way of steering a robot arm that first cancels out the arm's own complicated physics, so that what is left behaves like a simple, well-behaved toy. A real arm is messy to control: a joint near the shoulder has to fight gravity pulling on the whole limb, it gets shoved around as other joints swing, and a heavy elbow feels totally different from a light wrist. Instead of ignoring all that, the controller keeps a math model of the arm's weights, lengths, and motions, and at every instant it computes exactly the torque (the twisting push at each motor) needed to counter gravity and these tug-of-war effects. With those known forces pre-paid, each joint is freed to respond cleanly to your commands.

Once the physics is cancelled, the hard nonlinear arm collapses into something the controller can treat as a row of independent, identical units — push a joint and it simply accelerates, with no surprises and no cross-talk from its neighbours. The controller can then add a plain, gentle correction on top (typically a PID-style nudge that pushes harder the farther the joint is from where it should be) and get crisp, accurate tracking of a desired path. In short, computed-torque control spends a model of the robot to buy itself an easy control problem.

The catch is that it is only as good as its model: if the robot's true weight, friction, or a carried load differs from what the math assumes, the cancellation is imperfect and the leftover errors must be mopped up by the correction term or by adaptive methods. That is why it shines on well-characterised lab and industrial arms and is often paired with techniques that tune the model on the fly.