Start Here: From the First Hour to Expert Practice

English · 简体中文

This is not a list of articles to read from top to bottom. Work in a loop: understand → build → break → measure → explain → review. Time estimates help scheduling; only artifacts and gates support progression.

Your first 30 minutes

git clone https://github.com/Dld0621/Embodied-AI-Zero-to-Hero.git
cd Embodied-AI-Zero-to-Hero
python -m pip install numpy
python scripts/run_curriculum.py --validate
python scripts/run_curriculum.py --diagnose
python scripts/run_pipeline.py --run simulation-data

You should obtain a validated curriculum contract, an evidence self-diagnostic, and a synthetic Pipeline result. These outputs show that the entry path executes. They do not demonstrate mastery or real task performance.

The first-hour challenge

1. Locate the Pipeline inputs, stages, metrics, and output files.
2. Change one parameter, predict the effect, and rerun it.
3. Deliberately create one failure, such as a wrong shape, illegal range, or missing field.
4. Save the command, environment, seed, result, and failure explanation.
5. Write your first experiment card.

If you can copy a command but cannot explain its data flow, predict a change, or isolate a failure, start at M00.

Place yourself with evidence

Run:

python scripts/run_curriculum.py --diagnose

Ask one question for every module: Can I show the required artifact and let another person review the gate?

• Taking a class, reading a paper, or running code is not passing evidence.
• Without evidence, use not_started or in_progress; background does not grant an automatic pass.
• Experienced learners may skip explanations, but not assessment.

Choose the first emphasis for your background

Background	Likely strength	Common gap	Suggested entry
Complete beginner	Willingness to run and inspect Python	Environment, units, frames, experiment identity	M00
Mechanical/control	Kinematics, dynamics, and control intuition	Python, data contracts, learning evaluation	Diagnose M00, then focus on M05–M07
CS/ML	Programming, training, and model implementation	Frames, contact, closed-loop control, safety	M02–M04
Robotics practitioner	System integration, simulation, or hardware	Matched baselines, leakage, ablations, uncertainty	M05 and M11

Generate a route

python scripts/run_curriculum.py --list-goals
python scripts/run_curriculum.py \
  --plan full-stack-expert \
  --hours-per-week 8

Modules with existing reviewed evidence can be omitted for planning:

python scripts/run_curriculum.py \
  --plan full-stack-expert \
  --completed M00,M01 \
  --hours-per-week 8

--completed does not validate evidence. Initialize a formal record instead:

python scripts/run_curriculum.py \
  --init-progress learner/progress.json \
  --goal full-stack-expert \
  --learner your-name

A passed module requires existing evidence, a reviewer, and a review date. Audit the record with:

python scripts/run_curriculum.py --report-progress learner/progress.json

The module loop

1. Diagnose: resolve prerequisites and gaps.
2. Learn: complete the derivation or code yourself.
3. Build: produce the minimum module artifact.
4. Break: inject at least one failure, perturbation, or counterfactual.
5. Measure: retain raw metrics, configuration, logs, and failures.
6. Explain: distinguish data, model, interface, control, and task-definition failures.
7. Review: use the assessment standard to pass, repeat, or stop.

What beginner-to-expert means here

Level	Capability that must be demonstrated
L0	Another person reproduces your experiment; you repair a deliberate fault
L1	You derive and numerically verify the mathematics and catch frame/numerical traps
L2	You build a bounded sensing–state–control–simulation loop and isolate faults
L3	You build data, policy, RL, world-model, and planning baselines and separate offline from closed-loop quality
L4	You own one task family end to end, including perturbations, recovery, and stage failures
L5	You independently reproduce a baseline, test a falsifiable idea, run controlled comparisons, and survive independent review

Expert practice is not finishing every page or lowering a training loss. See the three capstones for graduation evidence.

When stuck

• Environment failure: use setup troubleshooting and preserve the first error.
• Unknown prerequisite: run python scripts/run_knowledge_map.py --path-to <node>.
• The system runs but the task fails: measure stage-resolved error before scaling or retraining.
• Hardware trial: first pass Sim-to-Real and safety gates; simulation never grants hardware authorization.