Whole-body control

Control software that coordinates every joint at once, so a robot keeps its balance while its arms and hands do work.

• Reaching or lifting shifts the robot's centre of mass and creates reaction forces; legs, hips and the other arm must compensate in the same instant.
• The controller recomputes all joint commands hundreds to a thousand times a second, ranking goals: stay upright first, then hand position, then posture.
• Classical methods solve a physics optimisation every cycle; newer ones train a neural network by reinforcement learning over millions of simulated trials.
• The hard part is the 'sim-to-real gap': friction, joint flex and sensor delays differ from the simulator, and small errors compound into a fall.