Whole-Body Inverse Kinematics

Solves for a robot's entire set of joint angles at once, so the hands, feet, and torso all reach their targets together.

Ordinary inverse kinematics usually handles just one arm: given a target end-effector pose, it solves for the joint angles along that one chain. Whole-body inverse kinematics scales this up to the entire robot — a humanoid, an arm-equipped quadruped, a dual-arm mobile base — satisfying several goals at once: where both hands should be, where both feet should land, which way the torso should face, keeping the center of mass over the support region, all while respecting joint limits and avoiding self-collision. With many joints and goals that often conflict, there's generally no closed-form solution; a common approach writes each goal as a 'task,' uses a Jacobian matrix to relate joint velocity to task error, and solves iteratively with weighted least squares or a quadratic program (this is differential inverse kinematics), resolving conflicts through task weights or priorities. It only computes joint angles and doesn't consider force or torque, which is what distinguishes it from whole-body control. In embodied AI it's commonly used to retarget human motion-capture data onto a humanoid, to drive whole-body teleoperation, and to generate reference motions for reinforcement learning.

ExampleGMR (General Motion Retargeting) maps human motion onto more than a dozen humanoid robots, including Unitree's G1 and Booster's T1: for every frame, a whole-body IK solver built on mink (a differential-IK library based on MuJoCo) aligns the robot's hands, feet, and pelvis to the corresponding human body parts while respecting joint limits, solving for the joint-angle sequence; the developers report it runs in real time on a CPU.

Also called

WBIK, Whole-Body IK

Related

Inverse Kinematics (IK) · Motion Retargeting · General Motion Retargeting · Whole-Body Control · Jacobian Pseudoinverse · Task Prioritization

Sources

mink: Python inverse kinematics based on MuJoCo (GitHub)
GMR: General Motion Retargeting (GitHub)

As of

2026-09