license: other
license_name: cern-ohl-s-2.0-apache-2.0-cc-by-4.0
license_link: LICENSE.md
language:
- en
pretty_name: GEMs — Gynoid Entity Models
tags:
- robotics
- humanoid
- open-hardware
- urdf
- bill-of-materials
- specification
configs:
- config_name: parts
data_files: hardware/bom/parts.csv
default: true
- config_name: ebom
data_files: hardware/bom/ebom.csv
- config_name: mbom
data_files: hardware/bom/mbom.csv
- config_name: avl
data_files: hardware/bom/avl.csv
- config_name: tasks
data_files: realtime_config/tasks.csv
- config_name: scheduler
data_files: realtime_config/scheduler.csv
GEMs — GYNOID ENTITY MODELS
Open-source designs of a physical robot body.
Documentation: https://plonemraz.github.io/vault/gems/
GEMs specifies a ~1.65 m humanoid body whose purpose is to acquire physical experience on behalf of a controller that lives wholly off it: the Main Controller, on an IPC, for which each GEMs unit is a set of sensors and effectors. It is a platform specification: it declares capability envelopes and the trade-offs between them, and deliberately leaves the operating point to whoever runs the body.
This is a paper design. Nothing here has been built.
The plain reason is money. There is no funding behind this work and no budget to buy parts with — the actuator modules §2.6 calls for, a pack of cells, a multi-threat armour layup, thirty joints' worth of anything, are simply beyond what one person can pay for. So nothing has been procured, nothing has been assembled, and no figure in this repository has been measured on hardware. What exists is a specification, a conformance protocol, executable models, and a component selection drawn from published sources.
Every number below is therefore a derivation or a citation, never a measurement. The models check their own arithmetic and hold the specification to it; what they cannot do is check either against a physical body. No amount of internal consistency substitutes for that, and the figures should be read as what a body of this description would require — not as what one has been observed to do.
Design envelopes
GEMs does not publish one configuration. Structure, actuation and energy sit on a single coupled loop — armour raises mass, mass raises actuator demand, demand raises battery mass, battery mass raises mass again — so fixing any one figure fixes the rest. What the specification publishes is the envelope and the trade-off, not a chosen point on it.
| Property | Declared envelope |
|---|---|
| Height | ~1.65 m — 1.75 m until 2026-09-26 |
| Mass | 76–160 kg, depending on endurance and armour coverage; the reference design sits at ~96 kg, 2 h with a dock |
| Free-running endurance | Hours, under a hard ceiling — past it no convergent design exists at any price |
| Deep-sleep endurance | ~years, bounded by battery self-discharge rather than by standby electronics |
| Protection | Handgun-calibre ballistic + stab + everyday impact, as one multi-layer package |
| Actuation | 80–90 Nm/kg peak over module mass at ~16:1; 3–5 kW/kg declared, under review. Peak power is limited by the source, not the actuators |
| Uplink | mmWave, up to ~8 Gbps, ~1 ms PHY latency at short range |
| Sensing | 16–67 Gbps raw aggregate; ≥2:1 on-body compression is mandatory, ~8:1 realistic |
| Compute | The Main Controller wholly off the body, on an IPC; on each unit, processing only, by four tiers that do not decide — the Real-time Control Layer, MCUs, DAQ and the OBC for link loss. Balance loop ≥500 Hz and reflex ≤10 ms must be on the unit |
| Audit surface | Signed sensor–actuator trace; a readable trace remains emittable at quiescent power while the body sleeps |
Every figure above is reproducible — python scripts/gems_budget.py --check recomputes them against the chapters that state them.
Several of these envelopes are bounded by something other than the obvious candidate. Peak power is capped by the battery rather than the actuators; sleep duration is capped by cell chemistry rather than by the standby circuit; tactile fidelity is capped by link bandwidth and fabrication density rather than by ambition. The specification states which constraint actually binds in each case.
What is where
This repository holds the source: hardware designs, the firmware and software that run on them, and the specifications they implement. Issues, pull requests and releases belong here.
The same tree is mirrored as the Hugging Face dataset plonemraz/GEMs, a snapshot of main after each push. There the tables load as subsets — parts, ebom, mbom and avl from hardware/bom/, tasks and scheduler from realtime_config/ — for example load_dataset("plonemraz/GEMs", "parts"). They are the design's tables, not measurements: every caution above holds for them too. The dataset's DOI, 10.57967/hf/11442, names the snapshot of commit e2bb849.
The documentation pages are built and published from plonemraz.github.io so that they share the site's navigation and theme. GitHub Pages is intentionally disabled on this repository — /vault/gems/ is served by the main site.
Earlier speculative material about GEMs lives under /vault/fiction/ and is not part of this repository.
Repository layout
One directory per group, and the body's three design disciplines nest under hardware/ rather than sitting at the root beside it.
| Path | Contents | Status |
|---|---|---|
spec/ |
Platform specification — the capability envelopes and their derivations | ✅ ten chapters |
plan/ |
What remains between the specification and a complete design: definition of done, open decisions, work packages, and the firmware and software work down to the task | ✅ v1 |
protocol/ |
Platform conformance protocol, and an assessment of this repository's simulated body against it | ✅ v0.1 draft + record |
hardware/ |
The body itself | ◐ kinematics declared |
hardware/mechanical/ |
Material and mechanism selection, sourced; CAD to follow | ◐ selection done |
hardware/electrical/ |
Actuator, power, compute and bus selection, sourced, with joint-by-joint sizing | ◐ selection done |
hardware/sim-model/ |
URDF generated from the kinematics, audited against it | ✅ |
hardware/design/ |
Concept art, then industrial design and expressive capability (plan ID) | ◐ concept art |
hardware/bom/ |
EBOM, MBOM and SBOM from one definition, checked against each other; approved manufacturer list, blanks where not yet verified | ◐ 206 parts, one priced |
firmware/ |
Code that runs directly on the microcontrollers, bare metal or RTOS: joint control, battery management, secure boot and attestation, low-power beacon — and the Real-time Control Layer's balance, reflex and supervisor modules if D-3 makes its computer a microcontroller | ◐ architecture; source awaits a target |
software/ |
Code that runs under Linux on the embedded computers: capture drivers, feature extraction and compression, sensor fusion, log synchronisation, link management | 🔜 no source yet |
reference/ |
Executable specifications the on-body code is checked against: audit log, agency tagging. They run on no target | ✅ with tests |
realtime_config/ |
Timing policy, apart from code: each task's deadline class (the requirement) and its platform, scheduling policy and latency bound (how it is met), checked against spec/ and the firmware architecture |
◐ requirement set; scheduling waits for D-2, D-3 |
scripts/ |
The coupled mass–energy–power loop, executable; checks the figures in spec/ |
✅ |
Directories marked 🔜 do not exist yet. They are named in advance so that the place a file belongs is never in question at the moment it is added.
Firmware and software sit at the root rather than under one heading, because they are different disciplines under one specification chapter: firmware runs directly on a microcontroller, software runs under an operating system, and the line between them is load-bearing enough to show in the layout. How late each task may be is a separate axis, held in realtime_config/ and never inferred from the folder a file is in. Terms follow standard usage — see the glossary, which also says where the vocabulary of the RSIL contract belongs.
Only firmware/ and software/ hold code that will run on the body. The Python under reference/ is executable specification, checked against by the on-body ports. The generators and checks under hardware/, protocol/, scripts/ and realtime_config/ produce and verify the design; none of it runs on the body, and the SBOM marks it excluded.
Roadmap
| Status | Item |
|---|---|
| ✅ | Capability envelopes derived and internally consistent |
| ✅ | Platform contract mapped — every external requirement has a named home in the design |
| ✅ | Audit surface specified: attestation, audit log, low-power beacon |
| ✅ | Licensing settled |
| ✅ | Budget model, and a check that holds spec/ to its own arithmetic |
| ✅ | Platform specification published in this repository |
| ✅ | Conformance protocol (v0.1 draft) |
| ✅ | Firmware architecture and the reflex budget split across stages |
| ✅ | Audit log: format, hash chain, batch signing, verifier |
| ✅ | Agency tagging by efference copy — reference model for the firmware port, with protocol §7.1 run against it |
| ✅ | Conformance assessment of the simulated body — 9 of 16 unmet, and named |
| ✅ | Kinematic configuration — DOF, arrangement, reach |
| ✅ | Simulation model — 31 DOF URDF with a coupled three-segment spine and two-tier joint limits, generated from the kinematics and audited against it |
| 🔜 | Firmware source (awaits a target board) |
| 🔜 | Feature extraction, fusion, link management |
| ✅ | Component selection, sourced — actuators, power, compute, bus, materials, reducers |
| ✅ | Actuator torque density reconciled with f_act — 80–90 Nm/kg peak over module mass, the basis declared, the trunk sized from measured human strength, the arms from a declared 15 kg payload; f_act split into its scaling and fixed parts |
| ✅ | Work plan to a complete design — decisions, packages, firmware and software task lists (plan/) |
| ◐ | EBOM, MBOM and SBOM — 206 part numbers, 60 routings, CycloneDX SBOM; sources and prices filled only where verified (hardware/bom/) |
| 🔜 | Mechanical CAD |
| 🔜 | Electrical schematics |
Related specifications
GEMs answers what the body consists of and what it can do. It is deliberately silent on how the signal that body acquires becomes structured information — that belongs to a separate, co-ranked specification.
| Specification | Scope | Status |
|---|---|---|
| RSIL — Relational Sensory Integration Loop | The information-processing loop between a body's sensors and its responses | Read |
| DIL — Data Integration Loop | The same relational structure for an agent with no body at all | Read |
Neither document claims the other's territory. The relation is complementary scopes, not upper and lower tiers.
Both links go to the author's site, which carries the full PDF of each. Both papers are also deposited in the Hugging Face dataset plonemraz/Plones_documents, DOI 10.57967/hf/11444.
Scope boundary
One division runs through everything here: the body grants capability and declares what it costs; the controller decides what to do with it. Nothing is cut at the level of the body for reasons that belong to whoever operates it, and where a value is left blank, the blank is deliberate — it marks a decision that is not the body's to make.
This repository is the body's side of that division, and only that side. It covers hardware, the firmware that runs it, the software that turns what the hardware acquires into information, and the protocols that carry signal between them. The controller is not specified here — not its design, not its reasoning, not how it will behave. It is addressed in the written introduction, not in this source tree.
Two conditions hold across everything in this repository. Nothing may call for physics that does not exist — laboratory work that is expensive, unscaled or not yet on the market is allowed; invented physics is not. And every capability has to earn its place by serving what the body is for: gathering physical experience. Each mechanism carries a mark saying how far it stands from something that can actually be built.
License
Three licenses, one per kind of work. Full routing in LICENSE.md.
| What | License |
|---|---|
| Hardware designs | CERN-OHL-S-2.0 — strongly reciprocal |
| Firmware and software | Apache-2.0 |
| Specifications and documentation | CC-BY-4.0 |
Hardware is reciprocal so that derivative designs stay open; software and documents are permissive so that the rest of the ecosystem can actually use them.
Status
The platform specification is published here in ten chapters, the conformance protocol in draft, and the parts of the stack that do not need a target board — the budget model, the audit log and the simulation model — are implemented and tested. Component selection is sourced and recorded; mechanical CAD and electrical schematics are not drawn.
No part of this has been built or physically validated, and the reason is the one at the top of this file: there is no money to buy the parts with. The work is complete as a paper design and untested as a body.