Whole-body control

Whole-body control is a robot-control framework that solves for joint torques across all degrees of freedom simultaneously against multiple competing objectives — end-effector position, balance, joint limits, collision avoidance — as a single constrained optimization at each timestep. Distinct from hierarchical control (one controller per subsystem) and from joint-PID (each joint tracks an independent reference).

Standard approaches: task-space inverse dynamics (TSID), model-predictive control (MPC) over a short horizon, and the OCS2 toolkit. Atlas, Digit, Apollo, and Figure 03 all use some form of MPC-based whole-body control; the differences are in horizon length, model fidelity, and how aggressively the controller constrains foot-vs-ground contact. The shift from earlier ZMP-style planners to whole-body MPC is the technical inflection that made modern humanoid balance robust and unlocked loco-manipulation.