Whole-body control

What it means

Whole-body control coordinates multiple parts of a robot while accounting for how their motions interact. A hand reaching toward an object can change the balance of a biped, so controlling the arm independently may be insufficient. The controller considers several objectives together, such as body posture, contact forces and the desired position of an end effector. [1, 2]

One common approach assigns priorities to tasks and solves constrained optimization problems. The momentum-control work by Herzog and colleagues uses hierarchical inverse dynamics with cascaded quadratic programs on a torque-controlled humanoid. Whole-body control is the coordination problem; that optimization method is one solution. [1]

How it applies to humanoids

Moving a heavy limb changes a humanoid's mass distribution and inertia. Zhang and colleagues combine a predictive planner with hierarchical optimization to account for those effects during walking. The planner selects motion and contact-force targets, while the tracking layer coordinates joint motion to follow them. [2]

The concept also applies to wheeled humanoids. Zafar and colleagues describe a dynamically balancing wheeled platform that must coordinate gaze, body pose, lifting and end-effector placement while respecting joint angle and torque limits. Whole-body control therefore does not require walking legs. [3]

Terminology and implementation traps

Whole-body does not promise that every joint is controlled in every experiment. Read which joints, tasks and contacts the particular implementation includes. A controller demonstrated for balancing and reaching does not by itself establish finger manipulation, navigation or task planning. [1, 3]

A predictive planner and a whole-body controller can occupy different layers of one system. The heavy-limb study explicitly combines model predictive control with hierarchical quadratic programming. Treating those labels as competing product categories would hide how that architecture works. [2]

Model errors, sensor noise and computation deadlines matter. A mathematically feasible command still has to run on physical hardware with imperfect measurements; the torque-controlled humanoid study addresses these practical difficulties explicitly. [1]

How to read it on humanoidversus.com

Read dof_total, sensors and max_joint_torque_nm as hardware context. None is a whole-body-control score. The ai_stack field may name a controller when the record's sources identify one, but an unspecified stack should remain unknown.

This entry has no live table or model-specific robot links. A robot's ability to stand, walk or move several limbs in a video is insufficient evidence for attributing a particular whole-body controller. Use a cited implementation paper to establish that connection.

Further reading

Herzog and colleagues cover hierarchical inverse dynamics on hardware. Zhang and colleagues explain a planner-and-tracker architecture for heavy limbs, while Zafar and colleagues extend coordinated control to balancing wheeled humanoids. [1–3]