Lesson 8 · 13 min
Everything in this course becomes a plan: four steps, each with what runs, the criterion to move up, and the reason to stop and go back. In this lesson you write your own ladder in a YAML file, check it with a script, and simulate the queue of steps with a failure in the middle. The safety procedure with the robot powered on is in course 4.
Start the lesson
Goal: write the ladder sim → sim2sim → suspended → floor with a pass criterion and a stop criterion at each step, and check it with a script.
Click each step. You only move up when the criterion is true.
If a step fails, go back to the previous one with what you learned: a new identified parameter, a new randomization range, a new criterion.
Run it, watch a step fail, and run it again. That's the ladder cycle.
Illustrative simulation in the browser. The failure on the "Suspended" step is on purpose, to show the return.
Copy the template to ~/fth/escada.yaml and swap in your robot and your criteria. conferir_escada.py requires the 4 steps in order, with onde, criterio (command → output), and when to stop.
# escada.yaml — the deployment ladder for your policy (FTH 2, lesson 8)
# Each step: what runs, the criterion to move up (command → expected output) and when to stop and go back.
politica: unitree_rl_lab G1 29 GDL velocity v0 (ONNX)
robo: Unitree G1 EDU (29 GDL)
degraus:
- nome: sim
onde: MuJoCo 3.16 + G1 do Menagerie, com as faixas de randomização do treino
criterio: "python3 kit/randomizar.py --seeds 20 --faixa treino → sobreviveu: 20/20"
parar_se: qualquer queda dentro da faixa do treino
- nome: sim2sim
onde: outro MJCF (unitree_mujoco) e o caminho de mensagens do robô (DDS, rt/lowcmd e rt/lowstate)
criterio: "python3 kit/portao_sim2sim.py --passo 0.001 --dds → PORTÃO: passou (3 de 3 etapas)"
parar_se: queda, velocidade fora de ±0,15 m/s, pernas no limite de torque ou política abaixo de 45 Hz
- nome: suspenso
onde: robô no pórtico, pés fora do chão, elástico ou cabo de segurança
criterio: "ensaio de bancada + sysid → kp identificado e atraso medido; a política mexe as pernas sem bater no limite"
parar_se: junta batendo no limite, vibração, temperatura subindo, atraso maior que o simulado
- nome: chão
onde: piso plano, área isolada, cabo frouxo no pórtico, pessoa no botão de parada
criterio: "teste no chão → 10 s parado + 10 s a 0,3 m/s sem queda, três vezes seguidas"
parar_se: tropeço, queda, qualquer pessoa dentro da áreaF=${FTH:-$HOME/fth}; cd $F/sim2real && . $F/.venv/bin/activate
[ -f $F/escada.yaml ] || cp kit/escada-modelo.yaml $F/escada.yaml
python3 kit/conferir_escada.py $F/escada.yaml; echo "saída: $?"F=${FTH:-$HOME/fth}; cd $F/sim2real && . $F/.venv/bin/activate
grep -v "parar_se: tropeço" kit/escada-modelo.yaml > $F/escada-incompleta.yaml
python3 kit/conferir_escada.py $F/escada-incompleta.yaml; echo "saída: $?"4 degraus, 4 critérios · escada ok saída: 0
Four steps, four criteria, exit 0.
ERRO: degrau 'chão' sem parar_se 4 degraus, 4 critérios · 1 problema(s) saída: 1
A step with no stop condition is the most expensive mistake: on the floor, it's what tells you when to press the button. The script fails it.
The plan all the way to the floor, verifiable.
Copy the template, swap in your robot, policy and criteria (use the outputs you captured yourself in lessons 2, 3 and 6), and check it.
Done when: python3 kit/conferir_escada.py ~/fth/escada.yaml; echo "saída: $?" → 4 degraus, 4 critérios · escada ok and saída: 0.
For each step, write in the YAML a stop reason that you would know how to measure (with the command or the instrument).
Done when: the 4 parar_se fields cite something measurable (fall, ms, %, N·m, temperature).
Generate your own study plan with the generator below and mark it on your calendar.
Done when: the downloaded plano.md lists the 8 lessons with the practice and the "done when" of each one.
Four main deliverables. Click a step to see its "done when".
The gap of a ready-made policy, measured, and the coverage of the training randomization compared with the outside world.
The "real robot" actuator identified on the gantry and the latency measured per step, with the compensation that makes the "real" one walk.
The simulator corrected with data and the policy passing through another model and through the manufacturer's DDS.
The twin and the proof on the physics, and the ladder with a criterion to go up and to stop at each step.
Four deliverables in sequence. The ladder is what course 4 (operating) applies to the real robot.
Order matters: the criterion for each step uses the output of an earlier lesson.
Pick a start date and how much time you have per week. The plan lists the practices for each lesson.
The explanation appears after your answer.
Question 1 of 3
On which step do you do the bench test, the sysid, and the latency measurement on the real robot?skill
Question 2 of 3
What makes a step's criterion verifiable?
Question 3 of 3
A step failed (real latency greater than simulated). What does the ladder tell you to do?
Pick one answer. The explanation appears after you try.