MUJOCO · HUMANOID · WHOLE-BODY CONTROL

Active Stance: Coordinated Whole-Body Control

Write a controller that commands a real MuJoCo humanoid into a deep, coordinated stance and holds it upright, the active counterpart to the passive support base.

01Challenge

Command the humanoid's leg joints to squat into a deep stance (pelvis below 0.48 m) and stay standing. Bend only the knee and it topples; you have to coordinate the ankle and hip so the center of mass stays over the feet. Your Python controller drives the real actuators.

02Model

The support base kept a leaning body upright with no controller. Now you add one. A humanoid's joints are driven by actuators that servo to the angles you command, so control here is choosing those angles well. The catch is coordination: drop into a squat by bending only the knees and the hips swing the mass backward off the feet, and the robot tips. Real bodies solve this by folding at three joints at once, the ankle dorsiflexing and the hip flexing forward as the knee bends, so the center of mass stays parked over the soles the whole way down. That is the whole-body coordination behind every stance, crouch, and the balance controllers that ride on top of it. You will command the three joints together and hold a deep stance without falling.

03Guided practice

1. What you return. control(obs) returns (hip, knee, ankle) target angles for both legs. 2. Go deeper. Make the knee more negative (deeper bend) to lower the pelvis below 0.48 m. 3. Coordinate or fall. A deep knee alone tips the robot. Flex the hip forward (larger positive hip) and dorsiflex the ankle (larger positive ankle) to keep the center of mass over the feet. 4. A stance that holds: around hip = 1.0, knee = -1.7, ankle = 0.85 reaches a deep, upright stance. Tune from there.

04Feedback

PASS WHEN PASS when the humanoid holds a deep stance (pelvis at or below 0.48 m) while staying upright: the ankle and hip coordinated with the knee to keep the center of mass over the feet.

If your run fails, check:

It tipped over: you likely bent the knee without compensating. Increase the hip flex and ankle dorsiflexion so the center of mass stays over the feet.
Not deep enough: the pelvis is still above 0.48 m. Bend the knee deeper (more negative) and coordinate the hip and ankle.
Define control(obs) returning three numbers: (hip, knee, ankle) in radians.

05Retrieve & space

Why does bending only the knee tip the humanoid? (the center of mass leaves the support base).
Which joints must move together for a balanced deep stance, and in which direction? (ankle dorsiflexes, hip flexes forward, as the knee bends).
What does a position actuator do with the angle you command? (servos the joint toward it).

06Mastery & project

Command a coordinated three-joint stance that stays balanced in real physics: the active whole-body control the passive support base leads into.

The whole-body coordination under every crouch, form, and balance controller; the full push-recovery and walking controllers extend it in the play/humanoid pilot.

GO FURTHER · THE LIVE PLAYGROUND

Pilot the full biped ↗ — drive the same MuJoCo humanoid yourself while the balance controller holds it up

← Standing Up: the Support Base Stepping to Recover: the Capture Point →

MUJOCO · HUMANOID · WHOLE-BODY CONTROL

Active Stance: Coordinated Whole-Body Control

Write a Python controller that commands the humanoid's hip, knee, and ankle. It runs on the real MuJoCo humanoid; hold a deep stance without tipping.

You read

obs each tick: pelvis height, upright measure, and time

You change

the hip, knee, and ankle targets your control(obs) returns

Fixed

the real MuJoCo humanoid, its actuators, and the balance check