E1 · Bench ready
10 minRun the setup block. If course 0 already set up ~/fth, it only adds what's missing.
Done when: the block ends with unitree_rl_lab 4960b84 (2025-11-19) and mujoco 3.16.0 · onnxruntime 1.31.0 · torch 2.14.1+cpu.
Lesson 1 · 16 min
Why locomotion became RL
The Unitree G1 in this course walks because a neural network sends 29 joint targets every 20 ms. In this lesson you set up the bench, run the ready-made Unitree G1 gait policy in MuJoCo, watch the same robot fall without it, and see why classical control gave way to RL in massive simulation.
Start the lesson
In this lesson you will
- 1Classical control (model + optimization) × policy learned in simulation
- 2The G1's open gait policy (unitree_rl_lab) walking in MuJoCo
- 3With policy: 6.6 m in 10 s. Without policy: fall in 1.4 s
- 4The ~/fth workbench with pinned versions
Goal: run the ready-made G1 gait policy in MuJoCo, measure what it does, and compare it with the robot without a policy.
01Two ways to make a humanoid walk
Click each block. The first three are the classical path; the last three are what the course practices.
Click each block in the diagram to see what it does.
MPC lives on as a concept and wherever a guarantee is required. The practice in this course is the right-hand column: simulation, policy and PD.
02What runs in this lesson
Click each stage.
1. The robot
The 29-degree-of-freedom Unitree G1 from MuJoCo Menagerie, pinned to commit 62ab00f (the same one as course 0). Position actuator with the policy's kp and kd.
2. The policy
Unitree publishes in unitree_rl_lab a G1 velocity policy already trained in Isaac Lab and exported to ONNX, with the deploy.yaml (joint order, kp and kd, default pose). Apache-2.0 license. The kit pins commit 4960b84.
Running in MuJoCo a policy trained in Isaac Lab already means crossing from one physics engine to another. Course 2 (sim2real) turns that into a gate.
3. The loop
Every 20 ms the policy reads the state and sends 29 targets; MuJoCo integrates at 500 Hz (2 ms step) with the PD holding each joint at its target.
andar_demo.py counts how many meters the pelvis walked and whether it fell (pelvis tilted more than 70° or below 0.3 m, the same criterion as the lesson 3 environment).
4. No policy
--sem-politica leaves only the PD holding the default pose. Nobody corrects the balance and the robot falls: what keeps the G1 upright while walking is the policy's closed loop.
03Set up the ~/fth bench
If you did course 0, the venv and Menagerie are already there: the block only adds what's missing (onnxruntime, CPU torch, the policy). Pinned versions, tested on Oct 19, 2026.
Terminal · set up ~/fth
F=${FTH:-$HOME/fth}; mkdir -p $F && cd $F
[ -d .venv ] || python3.12 -m venv .venv
. .venv/bin/activate && pip install -q mujoco==3.16.0 numpy pyyaml onnxruntime==1.31.0
pip install -q torch==2.14.1 --index-url https://download.pytorch.org/whl/cpu
[ -d locomocao ] || git clone -q --depth 1 https://github.com/inematds/formacao-fth-locomocao-v72 locomocao
[ -d mujoco_menagerie ] || git clone -q --filter=blob:none --sparse https://github.com/google-deepmind/mujoco_menagerie
(cd mujoco_menagerie && git sparse-checkout add unitree_g1 && git checkout -q 62ab00f)
[ -d unitree_rl_lab ] || git clone -q --filter=blob:none --sparse https://github.com/unitreerobotics/unitree_rl_lab
cd unitree_rl_lab && git sparse-checkout set deploy/robots/g1_29dof/config/policy/velocity
git checkout -q 4960b84 && git log -1 --format='unitree_rl_lab %h (%ad)' --date=short
python3 -c "import mujoco, onnxruntime, torch; print('mujoco', mujoco.__version__, '· onnxruntime', onnxruntime.__version__, '· torch', torch.__version__)"Example from the repository
unitree_rl_lab 4960b84 (2025-11-19) mujoco 3.16.0 · onnxruntime 1.31.0 · torch 2.14.1+cpu
Two lines and no errors. Pin the commits: a policy.onnx from another commit may have a different input order. torch comes from PyTorch's CPU wheels page (the course doesn't need a GPU).
ERRO: não achei /tmp/sem-bancada/unitree_rl_lab/deploy/robots/g1_29dof/config/policy/velocity/v0/exported/policy.onnx. Rode a preparação da aula 1. saída: 1
The scripts look for everything in ~/fth, or in the folder set by the FTH variable. If the setup didn't run (or ran in another folder), the error shows the path that was missing.
04The G1 walking, and the G1 without a policy
10 s asking for 0.7 m/s forward. It exits with 0 if it didn't fall, 1 if it fell.
Terminal · walk with the policy
F=${FTH:-$HOME/fth}; cd $F/locomocao && . $F/.venv/bin/activate
python3 kit/andar_demo.py --vx 0.7 --s 10; echo "saída: $?"Terminal · PD only, no policy
F=${FTH:-$HOME/fth}; cd $F/locomocao && . $F/.venv/bin/activate
python3 kit/andar_demo.py --vx 0.7 --s 10 --sem-politicaTerminal · ask for 1.0 m/s
F=${FTH:-$HOME/fth}; cd $F/locomocao && . $F/.venv/bin/activate
python3 kit/andar_demo.py --vx 1.0 --s 10 | tail -3Example from the repository
política: unitree_rl_lab velocity v0 (ONNX) · 480 → 29 · 50 Hz comando: vx = 0.7 m/s por 10 s · simulação a 500 Hz, política a 50 Hz andou: 6.6 m para a frente · desvio lateral: 0.5 m · velocidade média: 0.66 m/s altura final da pelve: 0.77 m caiu: não saída: 0
6.6 m in 10 s (0.66 m/s for the 0.70 requested) and the pelvis ends at 0.77 m. The 0.5 m lateral drift is normal: nothing corrects the heading, only the speed.
política: nenhuma (só PD na pose padrão) comando: vx = 0.7 m/s por 10 s · simulação a 500 Hz, política a 50 Hz andou: 0.8 m para a frente · desvio lateral: 0.0 m · velocidade média: 0.57 m/s altura final da pelve: 0.33 m caiu: sim, em t = 1.4 s
The robot falls at 1.4 s. The "average speed" of 0.57 m/s here is the body tipping forward, not walking: always read the speed together with caiu.
andou: 9.0 m para a frente · desvio lateral: 1.2 m · velocidade média: 0.90 m/s altura final da pelve: 0.76 m caiu: não
9.0 m in 10 s. 1.0 m/s is the forward-speed ceiling the policy was trained for (the unitree_rl_lab command range: −0.5 to 1.0 m/s).
05Practice
See the policy work before you open up the parts.
E2 · The G1 walks
3 minRun the policy at 0.7 m/s for 10 s and note how many meters the pelvis walked.
Done when: python3 kit/andar_demo.py --vx 0.7 --s 10; echo "saída: $?" → caiu: não and saída: 0.
E3 · The contrast
5 minRun with --sem-politica and write, in one sentence, what keeps the robot upright when it walks.
Done when: the output shows caiu: sim, em t = 1.4 s; your sentence says it's the policy's closed loop (PD alone doesn't correct balance).
06Lesson quiz
The explanation appears after your answer.
Question 1 of 3
What does the G1's gait policy deliver every 20 ms?skill
Correct
The policy decides joint targets at 50 Hz; each joint's PD, much faster, turns the error into torque. That is why both layers exist.Not that one
The policy decides joint targets at 50 Hz; each joint's PD, much faster, turns the error into torque. That is why both layers exist.Review: Two ways to make a humanoid walkQuestion 2 of 3
Without a policy, the robot "walked" at 0.57 m/s and fell at 1.4 s. How do you read that number?
Correct
The pelvis moved because the robot fell forward. Read every locomotion metric together with the fall; it's the same trap as the training curves (lesson 6).Not that one
The pelvis moved because the robot fell forward. Read every locomotion metric together with the fall; it's the same trap as the training curves (lesson 6).Review: Set up the ~/fth benchQuestion 3 of 3
What did Figure report about Helix 02, in January 2026?
Correct
This is Figure's own report (1 kHz, 10 M parameters). The course research found no independent MPC vs. RL comparison: that is why the course treats MPC as a concept and RL as practice.Not that one
This is Figure's own report (1 kHz, 10 M parameters). The course research found no independent MPC vs. RL comparison: that is why the course treats MPC as a concept and RL as practice.Review: Two ways to make a humanoid walkQuiz complete
You got {a} of {n} right. Get them all right and the lesson is marked as complete.Choose an option. The explanation appears after you try.
07In three sentences
ChangeHumanoid gait went from model + optimization to a policy trained in massive simulation.
LayersThe policy sends 29 targets at 50 Hz; each joint's PD produces the torque.
MeasureWith a policy, 6.6 m in 10 s; without it, a fall at 1.4 s.