Lesson 4 · 15 min
Knee and wrist don't call for the same motor. In this lesson you see the four actuator types that show up in humanoids and where each one tends to win, learn how the MJCF describes a motor, and measure, by simulation, how much torque the G1's knee produces to hold the robot in a squat.
Start the lesson
Goal: measure the torque of each joint of the G1 in a squat and compare it with the limit the MJCF declares for it.
The industry has not converged on a single actuator, but on a map of the body (humanoid.guide report, reported). Click each type.
Everything on this map comes from industry and supplier reports: it is a trend, not a specification.
Click each attribute.
The Menagerie's G1 uses position actuators: you send the target angle in ctrl and MuJoCo applies proportional control with kp=500 and critical damping (dampratio=1).
It's the same scheme as humanoid RL policies: the network sends joint targets and a PD on the robot turns that into torque (course 1 of the program).
The Menagerie's H1 uses torque actuators: ctrl is already the torque, from -200 to 200 N·m at the hip and -300 to 300 at the knee.
Without a controller of your own, the H1 collapses in the first second. The G1 stands on its own because the position actuators hold the "stand" pose.
On a position actuator, inheritrange="1" copies the joint's range into the command: the target never leaves the mechanical limit.
With a torque actuator, the command range is the torque range.
On the G1, the torque ceiling lives in the joint (actuatorfrcrange="-139 139" on the knee). Even if the proportional term asks for more, MuJoCo clips it there.
This is the number the script uses to calculate the percentage of the limit.
position kp="500" dampratio="1" inheritrange="1" · joint left_knee_joint actuatorfrcrange="-139 139"
torque_agachado.py starts from the "stand" pose, bends the knee (hip and ankle bend by half, so the foot stays flat), holds for 2 s, and reads the torque of each joint. Then it checks the largest one with inverse dynamics.
F=${FTH:-$HOME/fth}; cd $F/anatomia && . $F/.venv/bin/activate
python3 kit/torque_agachado.py --joelho 90F=${FTH:-$HOME/fth}; cd $F/anatomia && . $F/.venv/bin/activate
for k in 0 30 60 90 120; do
echo "joelho $k°: $(python3 kit/torque_agachado.py --joelho $k | grep -E '^joelhos|^ATENÇÃO' | tr '\n' ' ')"
donejoelho dobrado 90° · altura da pelve 0.59 m · 8 contatos com o chão junta torque (N·m) limite % do limite left_knee_joint 31 139 22% right_knee_joint 31 139 22% left_hip_pitch_joint 2 88 2% right_hip_pitch_joint 2 88 2% conferência mj_inverse (left_knee_joint): 31 N·m joelhos: esquerdo 31 N·m · direito 31 N·m maior torque: left_knee_joint 31 N·m
31 N·m in each knee, 22% of the model's ceiling. Simulating at rest and mj_inverse give the same value: that is the sign the pose is in equilibrium.
joelho 0°: joelhos: esquerdo 0 N·m · direito 0 N·m joelho 30°: joelhos: esquerdo 11 N·m · direito 11 N·m joelho 60°: joelhos: esquerdo 21 N·m · direito 21 N·m joelho 90°: joelhos: esquerdo 31 N·m · direito 31 N·m joelho 120°: ATENÇÃO: o robô caiu. joelhos: esquerdo 2 N·m · direito 2 N·m
Torque rises almost in a straight line up to 90°. At 120° the robot fell and sat down, and the "torque" reading collapses: a fallen robot's number is not a measurement. That is why the script warns you when the pelvis drops below 0.4 m.
Measure, sweep, and compare with the real robot.
Run the "knee at 90°" block and check that both paths (simulation at rest and mj_inverse) agree.
Done when: python3 kit/torque_agachado.py --joelho 90 ends with maior torque: left_knee_joint 31 N·m, with conferência mj_inverse (left_knee_joint): 31 N·m.
Run the sweep and write down the knee torque at each angle in a table. At what angle does the robot stop squatting and fall?
Done when: the sweep shows joelho 60°: joelhos: esquerdo 21 N·m · direito 21 N·m and the 120° line shows ATENÇÃO: o robô caiu.
Resellers cite a 90 N·m knee peak for the base G1 and 120 N·m for the EDU (reported). Calculate how much the 31 N·m of the squat represents in each case, and say whether there is margin left to jump.
Done when: you wrote down three percentages (22% of the model, about 34% of the base, and 26% of the EDU) and one sentence about margin, noting that the reseller numbers are reported.
The explanation appears after you answer.
Question 1 of 4
Why does the Menagerie G1 stand up with no controller of your own, and the H1 does not?skill
Question 2 of 4
Where does the harmonic reducer usually show up in a humanoid?
Question 3 of 4
Simulating at rest gave 31 N·m at the knee, and so did mj_inverse. What does that agreement show?
Question 4 of 4
At 120° the script read 2 N·m at the knee. What should you conclude?
Pick an option. The explanation shows up after you answer.