Open Humanoid Engineering Rev. 05

Open curriculum / Physical AI

Revision 05 · September 2026

From chemistry to whole-body intelligence

Build intelligence from matter up.

An open, AI-native curriculum for teams designing digital humanoids that can be explained, tested, challenged, and trusted.

A proposed education path, not an accredited degree or enrollment offer. The projects and laboratory architecture describe intended work; no completed physical humanoid is claimed.

Part of the aserdargun learning system
Explore the curriculum See the 2026 frontier
Exploded engineering schematic showing a complete humanoid and its mechanical assemblies
Concept illustration / System 00 Structure · Actuation · Sense · Power · Compute
Shared design target
01 humanoid
Foundational spine
05 disciplines
Open path
04 + 01 years
Team output
Public evidence

Matter-to-intelligence spine

  1. Chemistry
  2. Materials
  3. Mechanics
  4. Electrical
  5. Computer Science

Biomechanics + Neuroengineering

A system of responsibility, not a list of subjects.

The operating principles

Engineering has to become whole again.

Software made abstraction abundant. AI made generation abundant. Physical intelligence restores the hard constraints: matter, energy, time, uncertainty, people, and consequence.

  1. 01

    Matter before models.

    Intelligence begins with chemistry, surfaces, structures, friction, heat, and failure. A model is useful only when it remains accountable to the world.

  2. 02

    Evidence before claims.

    Every decision points to an equation, a simulation, a test, a result, and a stated limitation. Reproducibility is the native language of trust.

  3. 03

    AI as collaborator, never alibi.

    Teams use AI from day one, expose its assumptions, challenge its answers, inject faults, and retain responsibility for the final engineering judgment.

  4. 04

    Whole body, whole system.

    Perception, planning, control, mechanics, power, and human interaction are assessed together. Local success cannot hide system-level failure.

  5. 05

    Open progress compounds.

    Share reproducible laboratories, failures, evidence, and appeals after human review for privacy, permissions, and safety. Every verified contribution becomes a stronger starting point for the next team.

The field moved. The curriculum moved with it.

Primary source of truth · 2026 frontier signals

A humanoid is now an embodied agent stack.

The frontier is shifting from isolated robot skills toward long-horizon reasoning, multi-embodiment policies, on-device action, open simulation, collaboration, and explicit human intervention.

  1. 01
    Embodied reasoning

    Plan, act, observe, and recover across long tasks.

    Teach task decomposition, progress tracking, uncertainty, and escalation.

  2. 02
    Multi-embodiment policies

    Transfer intent across different bodies and tools.

    Separate high-level intent from embodiment-specific control and validation.

  3. 03
    Diagnostic simulation

    Measure generalization, not benchmark familiarity.

    Separate training and evaluation scenes; report trial counts, uncertainty, subtask progress, and failure modes.

  4. 04
    On-device autonomy

    Keep critical perception and action close to the robot.

    Design for latency, power, degraded networks, and bounded fallback behavior.

  5. 05
    Human-centered safety cases

    Know when to stop, refuse, or ask for help.

    Evaluate proximity, tool calls, uncertainty, and safe human intervention.

Primary sources

Google DeepMind · Whole-body intelligence · 30 Jul 2026 NVIDIA · Isaac GR00T 1.7 · 7 Jul 2026 NVIDIA Research · Policy evaluation · 11 Jul 2026

Sources reviewed . DeepMind informs the reasoning, transfer, on-device, and intervention signals; NVIDIA informs the policy workflow and diagnostic evaluation signals. These are research and vendor reports. The curriculum responses are our interpretation; the reports do not validate an ENG robot or establish general deployment readiness.

One integrated build, deepened through four public team cycles.

Primary source of truth · Four-year spine

One humanoid. Four years. Every course a project.

Each proposed year deepens one layer while carrying every previous layer forward. Begin with digital models; record separately any measurements from supervised hardware work.

  1. Year01

    Chemistry → Materials

    Intelligent limb

    Interface chemistryAI-designed materialsFailure mechanics

    Cells, surfaces, polymers, metals, composites, fatigue, and a virtual human-scale joint grounded in anatomy.

    Proposed team evidence · material passport + joint failure memo
  2. Year02

    Mechanics

    Fixed-base upper body

    Dexterous handsDigital twinSynthetic data

    Actuators, transmissions, arms, hands, tribology, manufacturability, grasping, and tool use in a circular workcell.

    Proposed team evidence · CAD + finite-element model + reproducible manipulation study
  3. Year03

    Electrical

    Walking full-body twin

    Edge controlPower + thermalFault injection

    Drives, battery management, sensing, real-time networks, balance control, and functional safety under injected faults.

    Proposed team evidence · simulated fault campaign + safe-stop envelope; supervised hardware-in-the-loop where available
  4. Year04

    Computer Science

    Autonomous digital humanoid

    Embodied reasoningVision-language-action policiesHuman intervention

    Whole-body control, robot learning, vision-language-action systems, planning, cybersecurity, intervention, and trustworthy autonomy.

    Proposed team evidence · held-out tasks + trial counts and uncertainty + autonomy safety case

Continuous human reference Biomechanics + Neuroengineering run through every year: anatomy, movement, motor control, ergonomics, injury limits, intent, and human–robot interaction.

The laboratory is software. Trust comes from its evidence.

Primary source of truth · AI lab pipeline

The laboratory is software.

The proposed laboratory architecture asks teams to reproduce, explain, extend, challenge, and publish AI-generated laboratories with their own AI agents. The packages, workspaces, evidence graph, and jury below are a design specification; this manifesto does not execute them.

  1. 01Reproduce

    Pin versions, inputs, and seeds.

  2. 02Explain

    State assumptions, units, and limits.

  3. 03Extend

    Change one factor; keep a baseline.

  4. 04Challenge

    Test held-out cases and injected faults.

  5. 05Publish

    Release evidence after human review.

System / 01

AI Lab Package

manifest.yaml
models/
instruments/
faults/
evidence.schema.json

Versioned physics, instruments, data, fault scenarios, and protected tests.

System / 02

Project Workspace

CAD · CODE · MODELS · RUNS · DECISIONS

Shared engineering work with transparent AI provenance and human ownership.

System / 03

Evidence Graph

CLAIM → MODEL → RUN → METRIC → LIMIT

Every claim remains connected to its result and known boundary. Label evidence as simulated, estimated, or measured; include units, trial counts, and uncertainty.

System / 04

AI Jury Report + Open Appeal

PASS / CONTESTED / INSUFFICIENT

AI reviewers would publish reasoning and dissent for human review. People authorize publication and resolve appeals; anyone may submit counter-evidence.

Breadth first. Precision after the system is understood.

The fifth-year frontier

Specialize at the intersections.

After four proposed shared years, a fifth-year frontier portfolio turns breadth into a precise research direction. These are portfolio pathways, not awarded qualifications.

  1. 01

    Computational Chemistry, Energy & Interfaces

  2. 02

    AI-Designed Materials & Soft Robotics

  3. 03

    Humanoid Mechanics & Dexterous Manipulation

  4. 04

    Intelligent Power, Actuation & Edge Systems

  5. 05

    Embodied AI & Robot Learning

  6. 06

    Neuroengineering & Human–Robot Symbiosis

  7. 07

    Trustworthy Digital Twins & Autonomous Circular Manufacturing

Year 04Integrated Humanoid Engineering Portfolio
Year 05Advanced Open Engineering Portfolio

Simulation earns the right to ask for physical validation.

Evidence required

If it survives the digital world, it earns the chance to enter the physical one.

A design that survives multiphysics simulation, randomized environments, injected failures, energy and thermal limits, cyberattacks, unseen tasks, and human proximity scenarios may be proposed for human review as a Production Candidate.

That designation does not certify a physical robot. It means the evidence is strong enough to seek funding, manufacturing review, and independent physical validation. Named human reviewers must record the decision, unresolved risks, and conditions for further testing; an AI jury cannot authorize physical operation.

Physics
Multiphysics limits + reality-gap statement
Autonomy
Unseen-task results + uncertainty behavior
Safety
Fault response + human intervention envelope
Provenance
Models, runs, decisions, and AI contribution history
Review
Human decision + unresolved risks + conditions for further testing
Inspect the open repository