E1 · Inside × outside
3 minRun the sampling for both ranges.
Done when: python3 kit/randomizar.py --seeds 20 → faixa treino sobreviveu: 20/20 and faixa fora sobreviveu: 8/20.
Lesson 2 · 16 min
Domain randomization: training for the varied world
If the simulator never gets the world right, you train the policy in many randomly drawn worlds. In this lesson you read the ranges Unitree itself randomized when training the policy, draw 20 worlds inside them and 20 outside, and measure the difference. The policy handles what it saw in training and fails just past it.
Start the lesson
In this lesson you will
- 1Where randomization lives in a real project: the Isaac Lab events
- 2The G1 policy's ranges: friction, mass, and push (and what was left out)
- 320/20 inside the range, 8/20 outside it
- 4Why randomizing everything to the max doesn't work either
Goal: read the randomization ranges of a real training run and measure the policy's survival inside and outside them.
01From randomized training to the out-of-range test
Click each block.
Click each block in the diagram to see what it does.
Randomization is not "more is better": it's covering the world where the robot will walk.
02What Unitree randomized (and what it didn't)
Read in velocity_env_cfg.py, unitree_rl_lab 4960b84 (Nov 19, 2025). Click each item.
1. Friction
physics_material event, at startup: static and dynamic friction between 0.3 and 1.0 on all of the robot's bodies.
In this lesson's out-of-range draw, the worlds with friction near 0.2 (2 and 5) fall; the floor of the training range is the floor the policy has seen.
2. Mass
add_base_mass event: adds −1 to +3 kg to the torso_link. A real payload (extra battery, camera, gripper) can go beyond that.
3. Push
push_robot event, every 5 s: changes the base velocity by up to ±0.5 m/s in x and y. The kit applies the equivalent as a 0.1 s force on the torso.
There is also noise in the observations (angular velocity, gravity, joints).
4. Outside
There is no randomization of kp, kd or latency (delay) in this task. In lesson 1, kp alone already changed the gait (0.75 m/s when asking for 0.50); latency (delay) matters under a push (lesson 4).
Other projects randomize gains and latency (delay); what matters is knowing what your training covered before you power on the robot.
03The kit's two ranges
treino.json copies Unitree's ranges; fora.json goes a bit beyond and adds kp and latency (delay).
File · kit/mundos/treino.json
{
"nome": "treino",
"descricao": "as faixas que a própria tarefa Unitree-G1-29dof-Velocity randomiza no treino (unitree_rl_lab 4960b84, velocity_env_cfg.py, classe EventCfg)",
"atrito": [0.3, 1.0],
"massa": [-1.0, 3.0],
"empurrao_vel": [-0.5, 0.5]
}File · kit/mundos/fora.json
{
"nome": "fora",
"descricao": "além do treino: atrito um pouco abaixo de 0.3, carga maior e kp e atraso, que o treino não randomizou",
"atrito": [
0.2,
0.6
],
"massa": [
2.0,
8.0
],
"kp": [
0.6,
1.0
],
"atraso_ms": [
0,
30
],
"empurrao_vel": [
-0.5,
0.5
]
}04Sample and count
randomizar.py samples 20 worlds per range with a fixed seed (7) and runs 10 s of walking in each one, with the push at t = 3 s.
Terminal · inside and outside the range
F=${FTH:-$HOME/fth}; cd $F/sim2real && . $F/.venv/bin/activate
python3 kit/randomizar.py --seeds 20Terminal · the first worlds outside the range, one by one
F=${FTH:-$HOME/fth}; cd $F/sim2real && . $F/.venv/bin/activate
python3 kit/randomizar.py --seeds 20 --faixa fora --detalhe | head -6Example from the repository
faixa treino sobreviveu: 20/20 faixa fora sobreviveu: 8/20
Inside the training range, 20 out of 20. Outside, 8 out of 20.
1 atrito=0.45 massa=7.38 kp=0.91 atraso_ms=6 empurrão 172 N → de pé 2 atrito=0.20 massa=6.93 kp=0.92 atraso_ms=14 empurrão 119 N → caiu 3 atrito=0.30 massa=4.67 kp=0.80 atraso_ms=16 empurrão 219 N → caiu 4 atrito=0.45 massa=7.93 kp=0.69 atraso_ms=4 empurrão 194 N → caiu 5 atrito=0.21 massa=5.09 kp=0.79 atraso_ms=28 empurrão 50 N → caiu 6 atrito=0.40 massa=3.49 kp=0.60 atraso_ms=6 empurrão 131 N → caiu
World 1 has a mass of 7.4 kg and stays up; world 4, at 7.9 kg, falls; world 6 has friction 0.40 and falls. No single parameter explains the score: outside the range, it is the combination. Looking at just one world and blaming one parameter is the common mistake; compare scores and, to claim a cause, vary one parameter at a time (E2).
05Practice
Measure what the training covered.
E2 · Only kp outside
5 minWrite a range identical to the training one, but with kp between 0.5 and 0.8 (the block below already does it), and run it.
Done when: python3 kit/randomizar.py --seeds 20 --faixa ~/fth/so-kp.json → faixa so-kp sobreviveu: 11/20.
E3 · What you would randomize
10 minFor a G1 that will walk on a factory floor with a 2 kg camera on its head, write in three lines what you would add to the training ranges and why.
Done when: three lines, each with a parameter, a range, and the measured reason (e.g., "kp 0.5–1.0: only kp outside gave 11/20").
06The E2 range
Same ranges as the training, plus kp between 0.5 and 0.8.
Terminal · only kp outside the range
F=${FTH:-$HOME/fth}; cd $F/sim2real && . $F/.venv/bin/activate
cat > $F/so-kp.json <<'J'
{"nome": "so-kp", "atrito": [0.3, 1.0], "massa": [-1.0, 3.0], "kp": [0.5, 0.8], "empurrao_vel": [-0.5, 0.5]}
J
python3 kit/randomizar.py --seeds 20 --faixa $F/so-kp.jsonExample from the repository
faixa so-kp sobreviveu: 11/20
Just adding kp, which the training did not vary, knocks down 9 out of 20.
07Lesson quiz
The explanation appears after your answer.
Question 1 of 4
The policy survived 20/20 inside the training ranges and 8/20 outside. What does that measure?skill
Correct
Randomization buys robustness inside the range. Outside it there is no guarantee; that is why you read what the training covered before going to the robot.Not this one
Randomization buys robustness inside the range. Outside it there is no guarantee; that is why you read what the training covered before going to the robot.Review: From randomized training to the out-of-range testQuestion 2 of 4
In this policy, what does the unitree_rl_lab task not randomize?
Correct
Friction, mass and push are in the EventCfg. Gains and latency (delay) are not; kp alone already changed the gait in lesson 1, and latency (delay) knocks it down under a push in lesson 4.Not this one
Friction, mass and push are in the EventCfg. Gains and latency (delay) are not; kp alone already changed the gait in lesson 1, and latency (delay) knocks it down under a push in lesson 4.Review: What Unitree randomized (and what it didn't)Question 3 of 4
Why not randomize everything across the widest possible range?
Correct
A range that is too wide costs performance and training time. The rule is to cover the world where the robot will work, with some margin, and measure.Not this one
A range that is too wide costs performance and training time. The rule is to cover the world where the robot will work, with some margin, and measure.Review: From randomized training to the out-of-range testQuestion 4 of 4
Why does randomizar.py use a fixed seed (7)?
Correct
With a fixed seed, 20/20 and 8/20 repeat on this stack (captured twice) and you can compare two policies on the same set of worlds. On another CPU, the last digit may vary.Not this one
With a fixed seed, 20/20 and 8/20 repeat on this stack (captured twice) and you can compare two policies on the same set of worlds. On another CPU, the last digit may vary.Review: The kit's two rangesQuiz complete
You got {a} of {n} right. Get them all right and the lesson is marked as done.Pick one answer. The explanation appears after you try.
08In three sentences
RangesG1 training randomizes friction 0.3–1.0, mass −1 to +3 kg, and pushes; it does not randomize kp or latency.
Score20/20 inside the range, 8/20 outside, 11/20 with only kp outside.
RuleRead what training covered and measure outside it before the robot.