E1 · Kit commands
5 minRun the four kit commands.
Done when: python3 kit/comando_wbc.py; echo "saída: $?" → executado: sim and saída: 0.
Lesson 7 · 15 min
When a VLA drives the whole body
A VLA sees the scene, understands the request, and proposes an action a few times per second. Turning that into safe movement of 29 joints, without falling, is the job of whole-body control. In this lesson you send high-level commands to the simulated G1 and see what it carries out and what it refuses before moving.
Start the lesson
In this lesson you will
- 1What the VLA hands down: targets, tokens or actions
- 2Decoupled WBC: legs by RL, arms by IK
- 3The lower layer checks before it executes
- 4The end-to-end VLA flow on the G1 (GR00T N1.7 + SONIC)
Goal: send high-level commands to a simulated whole-body controller and see which ones it executes and which it refuses.
01From request to movement
Click each block.
Click each block of the diagram to see what it does.
The VLA proposes; whole-body control checks and executes. Safety cannot depend only on the model above.
02How this is built in 2026
Click each case. Facts with sources; VLA details in vla-v72.
1. Figure
Three layers: slow S2, S1 emitting joint targets at 200 Hz, and a learned S0 at 1 kHz, which replaced 109,504 lines of C++ (01/27/2026).
S0 was trained with more than 1,000 h of retargeted human motion and sim-to-real RL: lessons 2 to 4 of this course at industrial scale.
2. NVIDIA
On 05/07/2026 the GR00T-WholeBodyControl repository published the end-to-end VLA flow on the G1: collect data through teleoperation, fine-tune GR00T N1.7, and deploy with SONIC as whole-body control.
SONIC v1.1 (07/23/2026) was trained with wrist pose augmentation precisely for teleop and for VLAs that use this control.
3. Decoupled
Before SONIC, GR00T N1.5 and N1.6 ran on the decoupled WBC: RL on the legs, IK on the arms. This is the split the kit imitates.
Advantage: simple and predictable. Limit: the torso and legs don't help the arm reach farther.
03Four high-level commands
The commands file plays the role of a VLA's output. comando_wbc.py checks each one, executes it in 1 s (ramp + closed-loop correction at 50 Hz), and says whether the hand arrived.
Terminal · run commands
F=${FTH:-$HOME/fth}; cd $F/corpo-inteiro && . $F/.venv/bin/activate
python3 kit/comando_wbc.py; echo "saída: $?"File · kit/comandos.json
{
"descricao": "Saída de alto nível, como a de um VLA a poucos Hz: o QUÊ fazer. Alvos da mão em metros, relativos à pelve (x para a frente, y para a esquerda, z para cima).",
"comandos": [
{"acao": "alcancar", "mao": "direita", "alvo_m": [0.30, -0.20, 0.15]},
{"acao": "girar_tronco", "graus": 20},
{"acao": "alcancar", "mao": "esquerda", "alvo_m": [0.30, 0.25, 0.20]},
{"acao": "voltar"}
]
}Example from the repository
comando 1 · alcancar direita (0.3, -0.2, 0.15) → executado: sim (mão a 1.1 cm do alvo) comando 2 · girar_tronco 20° → executado: sim comando 3 · alcancar esquerda (0.3, 0.25, 0.2) → executado: sim (mão a 1.0 cm do alvo) comando 4 · voltar → executado: sim robô de pé: sim · pelve a 0.79 m executado: sim saída: 0
Four commands, four executions, hand about 1 cm from each target, and the robot standing at the end.
comando 1 · alcancar direita (0.9, -0.2, 0.15) → recusado: alvo a 0.92 m do ombro; o braço alcança 0.38 m comando 2 · girar_tronco 200° → recusado: fora do limite da cintura (-150° a 150°) comando 3 · pular → recusado: ação desconhecida (conheço alcancar, girar_tronco, voltar) robô de pé: sim · pelve a 0.79 m executado: não · 3 comando(s) recusado(s) saída: 1
Target 92 cm from the shoulder (the arm reaches 38 cm to the wrist), waist at 200° (the limit is 150°), and an action that doesn't exist: all three are refused before any movement, and the robot stays still and standing.
04Practice
Run, refuse, write your own.
E2 · Bad commands
5 minRun the bad commands file and read the reason for each refusal.
Done when: python3 kit/comando_wbc.py --comandos kit/comandos-ruins.json | tail -1 → executado: não · 3 comando(s) recusado(s).
E3 · Your sequence
10 minWrite ~/fth/movimentos/meus-comandos.json with five commands (both hands, a torso turn, and return) and run it. If any is refused, the reason gives the distance to the shoulder: the arm reaches 0.38 m to the wrist.
Done when: python3 kit/comando_wbc.py --comandos ~/fth/movimentos/meus-comandos.json | tail -1 → executado: sim.
05Lesson quiz
The explanation appears after your answer.
Question 1 of 3
The VLA says to reach a point 92 cm from the shoulder. What should the lower layer do?skill
Correct
Checking before executing is what keeps a bad command from turning into motion. The VLA can be wrong; the low-level control is the last barrier before the joints.Not that one
Checking before executing is what keeps a bad command from turning into motion. The VLA can be wrong; the low-level control is the last barrier before the joints.Review: Four high-level commandsQuestion 2 of 3
In decoupled WBC, who takes care of the legs?
Correct
RL on the legs, IK on the arms. In the kit, the legs hold the model's pose through PD, in the place where the policy would go.Not that one
RL on the legs, IK on the arms. In the kit, the legs hold the model's pose through PD, in the place where the policy would go.Review: From request to movementQuestion 3 of 3
What does the May 7, 2026 GR00T-WholeBodyControl flow put together?
Correct
Collection by teleop (lesson 5), VLA fine-tuning (course vla-v72), and SONIC underneath (lesson 4): the layers of this course, connected.Not that one
Collection by teleop (lesson 5), VLA fine-tuning (course vla-v72), and SONIC underneath (lesson 4): the layers of this course, connected.Review: How this is built in 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.
06In three sentences
The what × the howThe VLA says the what; whole-body control decides the how.
ContractThe interface (targets, tokens) is what connects one to the other.
CheckThe lower layer refuses what can't be done, before moving.