E1 · Workbench ready
10 minRun the setup block. If you already had ~/fth from course 0, confirm that nothing was reinstalled besides this course's clone.
Done when: cd ~/fth/mujoco_menagerie && git log -1 --format=%h → 62ab00f.
Lesson 1 · 15 min
From walking to doing: whole-body control
Walking is just the start. To pick up a box, wave, or obey a VLA, the humanoid has to coordinate legs, torso, and arms at the same time, without falling. In this lesson you set up the course workbench and, in a simulation loop, count the three layers that make this happen: who decides, who coordinates the body, and who moves each joint.
Start the lesson
In this lesson you will
- 1The three layers: the VLA (S2 + S1) decides what, the S0 coordinates the body, the PD moves the motor
- 2Why each layer runs at a different frequency
- 3The 2025–26 paradigm: a policy that tracks human motion
- 4The workbench in ~/fth reused from course 0
Goal: set up the bench and name the three layers of a full-body humanoid (VLA, S0 and motor PD) with the frequency of each.
01Who decides, who coordinates, who moves
Click each block.
Click each block of the diagram to see what it does.
The names S2/S1/S0 follow Figure's Helix 02 (01/27/2026): the VLA is S2 + S1, whole-body control is the S0. The VLA itself is the subject of the vla-v72 course.
02What changed from 2024 to 2026
Click each step.
1. Before
Humanoid control was built from hand-written blocks: a step planner, the foot's trajectory in the air, the support leg, and balance via ZMP or MPC (model predictive control).
It works and it is interpretable, but every new skill needs new engineering. Lesson 6 of FTH course 0 showed the center of mass and the support polygon behind it.
2. 2025–26
The open path became: human motion capture or video → retargeting (GMR) → a motion tracking policy trained with PPO across thousands of simulations (BeyondMimic, GMT, SONIC) → sim2sim → robot.
One policy imitates any reference motion and serves as the base for teleoperation and for the VLA. Figure described the same thing in Helix 02: a 10M-parameter controller at 1 kHz, trained on more than 1.000 h of human motion, which replaced 109.504 lines of C++.
Caveat
MPC × RL: the research for this course found no independent quantitative comparison in 2026. MPC remains strong where a guarantee is required; RL is the dominant practice in open repositories.
3. Decoupled
A widely used middle ground: RL on the legs (walking and balancing) and IK on the arms (reaching). This is the decoupled WBC from the GR00T-WholeBodyControl repository, used under GR00T N1.5 and N1.6.
This course's kit follows that split: legs in the model's "stand" pose, arms via IK.
03Set up the workbench (reuses the one from course 0)
If you took course 0, the venv and the Menagerie already exist, and only this course's clone is new. If you didn't, the same block creates everything. The Menagerie is pinned to the tested commit.
Terminal · set up ~/fth
F=${FTH:-$HOME/fth}; mkdir -p $F && cd $F
[ -d .venv ] || python3.12 -m venv .venv
. .venv/bin/activate && pip install -q mujoco==3.16.0 numpy
[ -d corpo-inteiro ] || git clone -q --depth 1 https://github.com/inematds/formacao-fth-corpo-inteiro-v72 corpo-inteiro
[ -d mujoco_menagerie ] || git clone -q --filter=blob:none --sparse https://github.com/google-deepmind/mujoco_menagerie
cd mujoco_menagerie && git sparse-checkout add unitree_g1
git checkout -q 62ab00f && git log -1 --format='menagerie %h (%ad)' --date=short
python3 -c "import mujoco, numpy; print('mujoco', mujoco.__version__, '· numpy', numpy.__version__)"Example from the repository
menagerie 62ab00f (2026-10-15) mujoco 3.16.0 · numpy 2.5.4
Two lines and no errors. A different MuJoCo version may change the last digit of the numbers in the next lessons: pin the version.
04Count the layers in a 2-second loop
hierarquia.py runs the G1 standing for 2 s. Every 200 ms (task) it switches the right-hand target; every 20 ms (S0) it redoes the arm IK; every 2 ms it takes one physics step with the motors' PD.
Terminal · the three layers
F=${FTH:-$HOME/fth}; cd $F/corpo-inteiro && . $F/.venv/bin/activate
python3 kit/hierarquia.py; echo "saída: $?"Excerpt · kit/hierarquia.py
if i % k2 == 0: # task (S2+S1 in a VLA): decides "what" (here, a script of targets)
alvo = alvos[n2 % len(alvos)]
n2 += 1
if i % k1 == 0: # S0: whole-body control, decides "how" the body gets thereExample from the repository
simulado: 2 s · passo de física 2 ms tarefa (VLA: S2+S1) : 10 chamadas (5 Hz) S0 corpo inteiro (IK) : 100 chamadas (50 Hz) PD do motor + física : 1000 passos (500 Hz) mão direita a 0.7 cm do último alvo · pelve a 0.79 m · de pé: sim 3 camadas saída: 0
1000 PD steps for 100 S0 decisions and 10 task decisions. The hand ends less than 1 cm from the last target and the robot is still standing.
ERRO: S0 (30 Hz) e tarefa (5 Hz) precisam dividir o passo de 500 Hz. saída: 1
The bottom loop sets the clock: the upper layers run every N steps. 30 Hz does not fit into 500 Hz a whole number of times. Use divisors (50, 100, 250 Hz).
S0 corpo inteiro (IK) : 200 chamadas (100 Hz)
Twice as many body decisions in the same time. Helix 02 takes this to the extreme: the learned S0 runs at 1 kHz.
05Practice
Workbench, loop and frequencies.
E2 · The three layers
5 minRun hierarquia.py and note how many times each layer was called.
Done when: python3 kit/hierarquia.py; echo "saída: $?" → 3 camadas and saída: 0, with 1000 passos in the motor PD.
E3 · Faster S0
5 minRun with the S0 at 100 Hz and compare the hand's final distance to the target with the one at 50 Hz.
Done when: python3 kit/hierarquia.py --wbc-hz 100 | grep S0 → 200 chamadas (100 Hz).
06Lesson quiz
The explanation appears after your answer.
Question 1 of 4
On a full-body humanoid, which layer runs fastest?skill
Correct
The PD runs at 500 Hz–1 kHz because the motor must correct the angle before the body moves. The task changes slowly; running it at 1 kHz would only waste compute.Not that one
The PD runs at 500 Hz–1 kHz because the motor must correct the angle before the body moves. The task changes slowly; running it at 1 kHz would only waste compute.Review: Who decides, who coordinates, who movesQuestion 2 of 4
What is the "decoupled WBC" in GR00T-WholeBodyControl?
Correct
Legs by RL (walking, balancing) and arms by IK (reaching): it is the middle ground used under GR00T N1.5/N1.6. The kit mimics this split with the legs held still in the model's pose.Not that one
Legs by RL (walking, balancing) and arms by IK (reaching): it is the middle ground used under GR00T N1.5/N1.6. The kit mimics this split with the legs held still in the model's pose.Review: What changed from 2024 to 2026Question 3 of 4
Why does hierarquia.py refuse the S0 at 30 Hz?
Correct
The 2 ms step sets the clock; the upper layers run "every k steps". At 30 Hz, k = 16.7: it doesn't exist. Real robots pick divisors for the same reason.Not that one
The 2 ms step sets the clock; the upper layers run "every k steps". At 30 Hz, k = 16.7: it doesn't exist. Real robots pick divisors for the same reason.Review: Set up the workbench (reuses the one from course 0)Question 4 of 4
What did Figure swap out in Helix 02, according to the course research?
Correct
Helix 02's S0 is learned: 10M parameters, 1 kHz, trained on more than 1,000 h of retargeted human motion, in place of 109,504 lines of C++ (01/27/2026).Not that one
Helix 02's S0 is learned: 10M parameters, 1 kHz, trained on more than 1,000 h of retargeted human motion, in place of 109,504 lines of C++ (01/27/2026).Review: What changed from 2024 to 2026Quiz complete
You got {a} of {n} right. If you get them all right, the lesson is marked as complete.Choose an option. The explanation appears after you try.
07In three sentences
Three layersThe VLA (S2 + S1) says what, the S0 coordinates the body, the PD moves the motor.
FrequenciesFrom slow to fast; the upper ones run every N steps of the one below.
2026The S0 became a policy that tracks human motion: that is the rest of the course.