Embodied AI · Learning and engineering

Understand the loop.
Build the system.
Prove the result.

A bilingual curriculum that connects first principles, runnable Pipelines, research routes, and explicit evidence boundaries.

New here? Start now Open the curriculum Map prerequisites Explore Pipelines 中文

45Knowledge nodes

14Foundation lessons

12Mastery modules

11Engineering Pipelines

7Research routes

3Capstones

Choose by outcome¶

START

Complete the first-hour challenge

Place yourself with evidence, generate a route, and retain your first experiment card.

LEARN

Build the prerequisite chain

Follow six stages from experiment discipline and mathematics to task systems and deployment evidence.

PREPARE

Configure a reproducible workstation

Select a reviewed stack, isolate dependencies, run layered checks, and retain an environment receipt.

BUILD

Execute one complete system

Start from inputs, retain intermediate artifacts, measure outcomes, and localize failures by stage.

RESEARCH

Turn a question into evidence

Choose a baseline, freeze a protocol, plan an ablation, and declare the promotion gate.

SPECIALIZE

Enter VLA and WAM from zero

Understand each algorithm family, select by constraints, build matched baselines, and progress to research-grade experiments.

One closed loop¶

01ObserveSensors, frames, state, uncertainty

→

02DecideGoals, policy, reasoning, prediction

→

03ActControl, limits, watchdogs, stops

→

04EvaluateTask metrics, failures, learning

Every method must eventually connect observation to evaluated action. A plausible command, open-loop loss, or completed script is not task success by itself.

Seven research routes¶

01Foundation Models and VLAPolicy · adapter · ablation 02Manipulation and ImitationBaseline · failures · closed loop 03Dexterity and TeleoperationRetargeting · contact · retention 04Navigation and AgentsState · planning · recovery 05Humanoids and LocomotionMotion · robustness · safety 06Perception and World ModelsUncertainty · predictive rollout 07Simulation, Data, and EvaluationDatasheet · benchmark · gate

Pipeline coverage¶

Smoke-tested · 8 Interface-tested · 2 Hardware-dependent · 1

SMOKESimulation and dataTask → expert → episodes → QA SMOKEVLA policyMultimodal data → policy → evaluation SMOKEWorld modelsTransitions → rollout → planning SMOKERL post-trainingMDP → PPO → evaluation → regression INTERFACERFM and cross-embodimentCanonical schema → adapter → safety INTERFACEEmbodied reasoningInstruction → skills → feedback → replan DOCSim-to-RealHIL → shadow → guarded rollout SMOKEDexterous retargetingLandmarks → geometry → optimization SMOKEPerception and stateCalibration → synchronization → fusion SMOKENavigation and locomotionState → planning → control → recovery SMOKEDexterous manipulationPre-grasp → contact → lift → hold

Evidence before claims¶

Execution is not performance. Simulation is not hardware validation.

The repository separates import, smoke execution, deterministic tests, benchmark evidence, and bounded hardware validation. A lower level never implies a higher one.

Validation policy → · Accuracy gate → · Benchmark report →

Verify locally¶

python scripts/run_knowledge_map.py --validate
python scripts/run_pipeline.py --validate
python scripts/run_learning_path.py --validate
python scripts/run_curriculum.py --validate
python scripts/audit_repository.py
python -m pytest tests/ -q

The root English README and 中文 README provide the repository overview. This site is the detailed learning and engineering layer.

Understand the loop.Build the system.Prove the result.