Self-Replicating Humanoid Robots
This is the most consequential technology question in robotics: can a robot build copies of itself? Not in theory — in practice, today, with real parts, real costs, real supply chains. The answer determines whether humanoid robots remain expensive bespoke machines or become self-propagating infrastructure that gets cheaper with every generation.
The Replication Spectrum
Five classes of self-replication, ranked from hardest (Class 1 — nobody has achieved it) to emerging (Class 5 — collective replication). A system's class is the highest it can honestly claim. Most real systems sit in Class 3-5.
Autotrophic
System produces every part of a copy of itself from raw matter — ore, regolith, CO2, sunlight. Zero imports.
None exist. Theoretical only.
- von Neumann universal constructor (1948)
- NASA lunar factory study (1980)
- Freitas & Merkle (2004)
Autotrophic-Lite
System produces all structural parts and most non-specialist subsystems from raw or near-raw feedstock. Imports only high-complexity items (ICs, rare earths, specialist chemicals).
None shipping. Aspirational only.
- MIT CBA Fab Lab 2.0
- ESA/NASA lunar ISRU concepts
- Adrian Bowyer's long-range RepRap vision
Partial Kinematic
System reproduces a defined subset of its own parts from standardised feedstock (filament, wire, resin). Imports the rest from a normal supply chain. This is the RepRap frontier.
Real and shipping. The honest ceiling for today.
- RepRap family (Darwin, Mendel, Prusa i3)
- Voron 2.4 / Trident
- Hod Lipson modular robot (Cornell 2005)
Assembly-Only
System takes already-manufactured parts and assembles them into a working copy of itself. No parts are made, only combined.
Most 'self-assembling' marketing claims live here.
- Pick-and-place + screwdriver demos
- Tesla factory rhetoric
- Most commercial assembly robots
Swarm / Distributed
No single node self-replicates; the collective, taken as one aggregate, produces copies of nodes.
Emerging. No honest commercial example yet.
- Fab-lab networks (MIT CBA)
- Warehouse-scale 'machine shops of machine shops'
The Key Insight: P% ≠ M% ≠ C%
A single number is a lie. Self-replication must be measured along three independent dimensions — and a system scoring high on one and low on others is the interesting case. This is Neumann's signature contribution to the field.
Proportion of discrete parts in the BoM that the machine can produce itself. Cheap to compute, easy to game — inflate with printed brackets and the number soars.
Favours designs with many small printed parts.
Proportion of the finished copy's mass that the machine produces itself. Favours machines that print their frames, penalises those that print only trim.
Usually much higher than P% for plastic-framed machines.
Proportion of the BoM cost the machine can cancel — valued at arm's-length market prices. The honest measure. Printed plastic is cheap; bought ICs and motors are expensive.
Usually the lowest of the three. The one that matters most.
Why the distinction matters — Prusa MK4 vs Original RepRap
| Dimension | Prusa MK4 | RepRap Mendel | What this tells you |
|---|---|---|---|
| P% | 25-35% | 40-55% | RepRap prints more of its own parts — its frame is mostly printed plastic |
| M% | 6-12% | 10-20% | Both still import most of their mass (metal, motors, electronics) |
| C% | 2-5% | 3-8% | Both are <=8% self-replicating by cost — printed plastic is cheap |
All values are estimates (est) with bounds from Neumann session-1 research. Primary-source BoM reconstruction is in progress. Calling a Prusa "self-replicating" without qualification is misleading — it is ≥95% imported by cost.
Products Relevant to Self-Replication
Only products that ship today. Roadmap items and vapourware are excluded — Neumann's rules of honesty prohibit scoring anything that doesn't have shipping hardware.
RepRap Darwin (2006-2007)
Open Source (RepRap Project, Adrian Bowyer)
All structural plastic vertices, corner brackets, axis carriages, extruder body, idler housings — the original printable-frame demonstration
Threaded rods, smooth rods, NEMA 17 steppers, Sanguinololu/RAMPS-era electronics, hotend, belts, bearings, fasteners
The original. Darwin is the canonical Class-3 reference machine — the first time a 3D printer produced parts that became another 3D printer. Brittle, slow, fussy by modern standards; numbers are estimates because no published primary BoM survives in tidy form. Bowyer's vision is the philosophical anchor for everything below.
Prusa i3 MK4/MK4S
Prusa Research
Plastic brackets, frame endcaps, fan ducts, cable guides, X/Y/Z carriages
Aluminium extrusions, linear rails, NEMA 17 motors, heated bed, PSU, control board, display, belts, pulleys, fasteners, hotend
The canonical Class-3 product. Prusa's print-farm is the world's largest ongoing Class-3 operation — but the printer is >=95% imported by cost.
Original RepRap Mendel
Open Source (RepRap Project)
All plastic vertices, carriages, idlers — frame topology is >=50% printed plastic by count
Threaded rods, motors, electronics, hotend, belts, bearings
Purer Class-3 than any commercial product. But low throughput, fussy build, obsolete electronics. The 2009 successor to Darwin and the high-water mark of pure-printed-frame design.
Voron 2.4 / Trident
Open Source (Voron Design)
All printed structural brackets, idlers, fan housings, cable chain mounts (70-90 printed parts)
Aluminium extrusions, linear rails, NEMA motors, BTT control boards, heater, bed, PSU, belts, hotend
High-performance Class-3 with an active community. Self-sourced BoM means transparency, but cost fraction is still low.
Lipson Modular Cube Robot (Cornell, 2005)
Cornell Creative Machines Lab (Hod Lipson et al.)
Replication unit of currency is a 10cm swivel-cube module. The robot assembled identical modules from a stock of pre-made cubes into a copy of itself — the canonical kinematic self-replication demo at the module level.
Every cube interior: motors, controllers, electromagnets, batteries, bearings, fasteners — i.e. the modules themselves are not made by the robot
Famous Nature 2005 paper. Honest about its scope: replication is at the level of pre-fabricated modules, not raw matter. Important historical proof that self-assembly into a like-shaped successor is achievable with then-current tech. Sets the academic ceiling above which industrial Class-3 must rise.
Open Source Ecology — LifeTrac VI
Open Source Ecology (Marcin Jakubowski)
Welded steel frame, hydraulic quick-connect couplers, modular power-cube interface, wheel hubs and brackets — all designed for fabrication in a community workshop with a CNC torch table, lathe, and welder
Hydraulic pump and motors, diesel/electric power-cube engine, controller electronics, tyres, bearings, fasteners
Distributed Class-3 — closer to Class-5 in spirit. The GVCS thesis is that 50 open-blueprints machines (tractor, brick press, CNC torch, induction furnace, 3D printer, ...) collectively self-replicate as an ecosystem rather than as individuals. Numbers are estimates: real builds vary 2x by builder skill. The political project is as important as the technical one.
Snapmaker 2.0 / Artisan
Snapmaker
Plastic shrouds, fan ducts, cable guides, replacement nozzle holders. With the laser/CNC modules attached the same machine can also engrave its own panels and rough-machine softer brackets in aluminium.
Closed proprietary linear-rail modules (the load-bearing axes are sealed motion components), NEMA steppers, controller PCBs, touchscreen, PSU, laser diode, CNC spindle, bed, fasteners
Lower P% / M% than Voron because the structural axes are closed proprietary modules — you cannot print a Snapmaker linear module on a Snapmaker. The compensating story is multi-process: one machine that can print, laser-cut, and CNC-mill is more useful per dollar to a builder. Class-3 in scope, Class-4 in temperament.
OpenROV Trident
OpenROV (open-source community)
Printed hull halves, thruster ducts, camera mount, internal frame brackets — entire body shell is printable in PETG or ABS
Brushless thrusters and ESCs, BeagleBone-class compute, camera module, tether, batteries, depth sensor, O-rings
Open-design more than self-fabricating, but the printable hull moves the M% above most desktop printers. Useful as a non-humanoid datapoint: open-blueprints + printable shell can deliver a working vehicle for the price of a high-end printer. Adjacent rather than central to humanoid replication.
Asimov v1
Menlo Research (open-source release)
MJF PA12 nylon shells, 7075 aluminium machined frame parts (open CAD); design + simulation files released for in-house printing/machining
25 actuators (ENCOS-class motors), Jetson-class compute, IMU, F/T sensors, batteries, fasteners, wiring harness
First humanoid in this index. Mechanical design is open enough to qualify as Class-3 in structure — but motors, electronics and batteries remain bought-in. Class drops to 4 if you outsource the MJF print + 7075 machining (the realistic case for most builders). Tracks our /companies/asimov-menlo profile.
Mini 6-DOF Open Arm
@IlirAliu_ (open-source GitHub)
Split-ring planetary gearboxes, inverted-belt differential wrist, all structural housings (3D-printed); custom STM32 firmware in C
Stepper motors, TMC5150 drivers, AS5048a encoders, BTT Octopus board, bearings, belts, fasteners
Highest P%/M% of any product in the index. Printed planetary gearboxes are the headline self-replication move — the part of the BOM that traditionally requires a dedicated supply chain is here printed in PLA/PETG. Doesn't change the stepper/encoder/driver gap, but proves the gearbox can be in-source for hobbyist tolerances.
LeSlider (SO101 + linear track)
@pham-tuan-binh (open-source GitHub)
All track mounts, pinion/track parts, end-stops, motor brackets (printed); shares SO101 motor bus so no extra controller
STS3215 servo (extra motor), V-slot extrusion or printed track teeth, fasteners, wiring
An extension, not a base machine. Demonstrates the compounding effect: when the host platform (SO101) already has an open BOM, a community member can extend it with predominantly-printed parts and an arbitrary-length track. The kit-builder analogue for HUMA: every base kit should be extensible by the same printable-bracket logic.
Pollen Robotics Reachy 2
Pollen Robotics (Hugging Face)
Printed shoulder shrouds, forearm shells, gripper fingers, head housing, cable raceways — a high fraction of the visible structure is printable in PA12 or PETG-CF on a hobby-grade printer
Dynamixel-class smart servos (the dominant cost), torso frame extrusions, mobile base wheels and motors, RealSense cameras, on-board PC, batteries, IMU, fasteners
Most credible open humanoid platform after the SO-arm family. Class-3 if you print the shells yourself; drops to Class-4 if you buy them pre-printed. Servo bill is the dominant unmovable cost — same gap as every other humanoid. Hugging Face acquisition (2024) suggests the platform will widen rather than close.
Tesla Optimus (Gen 2/Gen 3)
Tesla
In-house: 4680 battery cells, FSD-derived inference chip, custom rotary and linear actuators, hand mechanisms — all manufactured at Tesla factories, not by the robot itself
External suppliers for fasteners, sensors, harnesses, structural metals, rare-earth magnets, plus the entire industrial tool-chain (machining, casting, winding) that the robot does not perform
Vertical integration is not self-replication. Tesla designs and manufactures most of Optimus's parts inside its own factories — but Optimus does not. The robot is assembly-only at best, and even that is rhetorical: current units are hand-finished by humans on the line. Class-4 is the honest ceiling regardless of what the marketing says.
Unitree G1
Unitree Robotics
Unitree designs and produces its own M107-series joint motors, custom reducers, and control boards in-house — but at factory scale, not by the robot
NVIDIA Jetson Orin compute, 3D LiDAR, depth cameras, batteries, fasteners, structural alloys, rare-earth magnets, sensors
Largest global humanoid production volume (~1,200 units in 2024). Like Tesla, vertical integration is mistaken for self-replication. The G1 is interesting precisely because its low price makes it the ideal candidate for the Opportunity #5 teardown — buy one, disassemble, publish the BoM. No P%/M%/C% scores until that teardown happens.
K-Scale K-Bot
K-Scale Labs
Open CAD: shell parts could be printed by a builder, but the shipping kit arrives with parts already CNC'd / injection-moulded. Open K-OS Rust stack is the headline self-buildable layer.
Brushless actuators, bearings, harnesses, compute modules, batteries, sensors, fasteners — purchased pre-fabricated even when blueprints are open
Open-hardware in design language, Class-4 in physical reality. The kit's significance is the price point: $8,999 puts a complete humanoid below the cost of a Voltera NOVA. If the open CAD attracts a community that prints its own shells, K-Bot could shift toward Class-3 — but as shipped, no parts are self-fabricated by the robot or its owner.
Bambu Lab X1 Carbon
Bambu Lab
A handful of cosmetic clip-on parts — some users print replacement fan shrouds and cable clips, and that is the full extent of meaningful self-production
Injection-moulded enclosure, sealed CoreXY motion assembly, proprietary control board and firmware, AMS multi-material unit, hotend, sensors, PSU, lidar — all closed and not user-fabricable
The market leader. Bambu has explicitly chosen the opposite of RepRap: maximum manufacturability at scale, minimum self-fabrication. Excellent printer, terrible self-replicator — and the industry's best evidence that commercial throughput and Class-3 properties are in revealed tension. Included here to make the contrast visible.
Anycubic Kobra series
Anycubic
Cosmetic spool holders, replacement fan ducts, optional cable clips — anything beyond is non-trivial because the structural injection-moulded body and proprietary motion assembly are not user-fabricable
Steel/aluminium frame, motion components, proprietary mainboard, touchscreen, hotend, bed, PSU, harnesses, fasteners — almost the entire BOM
Genealogically a great-grandchild of RepRap (i3 lineage), commercially in the Bambu camp: closed firmware, closed control board, sealed hotend, no published BOM. Volume leader at the budget end. The Kobra is what mass-market 3D printing looks like when the open-design covenant is dropped — strong argument for why the RepRap aftermarket exists.
Nano Dimension DragonFly IV
Nano Dimension
Multi-layer PCBs (up to 10 layers), embedded conductors + dielectrics, some passives
Active components (ICs, transistors, discrete semis), high-frequency connectors
If any single product closes the electronics gap, this is it. Class-2 of the electronics category is not yet — but the DragonFly is the only commercial primitive in striking distance.
Voltera NOVA / V-One
Voltera
Single-side and flex circuits, conductor traces via conductive ink
Multi-layer boards at density, ICs, all active components
Desktop conductive-ink PCB prototyping. Accessible entry point for electronics self-fabrication, but limited to simple circuits.
Desktop Metal Shop System
Desktop Metal (Stratasys)
Metal structural parts (brackets, small hardware) with ~1% dimensional tolerance
Precision bearings, motor laminations, fasteners at commodity cost
Bound-metal binder jetting for shop-scale steel/copper/stainless. Shrinks the 'imported metal parts' list for a theoretical Class-2 replicator.
eFlesh — magnetic tactile sensor
Open-source research (lukas_m_ziegler / paper authors)
Magnetic-flux-transducer pad printable in elastomer + Hall-sensor PCB (single-sided)
Hall-effect IC, neodymium magnets, microcontroller, wiring
Component-level proof point, not a machine. The interesting move: a tactile sensor — the part of every humanoid's BOM that historically required custom MEMS or commercial OEM modules — is now hobbyist-fabricable for $5. Direct entry on the Class-2-of-electronics roadmap (next to Voltera/DragonFly). If repeated across F/T, IMU, and joint encoders, the sensor sub-budget collapses.
Notable: The Anti-Replication Trend
Market leaders by unit volume (Bambu Lab X1 Carbon, UltiMaker S5, Creality K1) score less than 5% on P%, less than 2% on M%, and less than 1% on C%. They have deliberately de-self-replicated their designs to maximise manufacturability at scale. Commercial viability and Class-3 properties are in tension — a revealed preference the industry should respect.
Adjacent: The Marlin / Klipper Firmware Layer
Not a product on the index — a primitive every Class-3 machine depends on. Marlin (since 2011, GPL) and Klipper (since 2016) are the firmware substrate that runs Prusa, Voron, RepRap, and most descendants. The firmware itself is freely copied and recompiled by every builder, making it a perfect P%/M%/C% = 100% layer in software. Self-replication is bottlenecked by atoms, not bits — and Marlin/Klipper prove the bit layer is already fully open.
Kit SR — The Self-Replication Kit
HUMA currently offers three kit humanoid robot templates. We are researching a potential fifth variant — Kit SR — designed from Neumann's findings to maximise the proportion of parts the robot can manufacture itself.
Maximise C% — the cost fraction the machine can self-produce. P% is easy to game. C% is honest.
Electronics and actuation. Printed plastic is solved. Printable motors and desktop PCB fabrication are the frontier.
Neumann's taxonomy, RepRap heritage, gap analysis of structural vs electronics vs actuation categories.
Coming when the research is ready — we don't oversell. Neumann is actively scoring products, reconstructing BoMs, and mapping the gap between today's Class-3 ceiling and a credible Class-2 kit. This page will update as findings mature.
Commercial Opportunities
Self-replication research creates real business opportunities across three time horizons: harvest (0-2 years, RepRap economy), bridge (2-7 years, gap-closing tooling), and bet (7-20 years, genuine Class-2 capability).
The Self-Replication Index
0-6 monthsNobody has a consistent, cited, numeric measure of 'how self-replicating is this machine?' Publish an annual index with ~50 machines scored using P%/M%/C% + class.
Industry-standard reference. Inbound from founders, journalists, investors. Conference talk circuit. Licensing to consultancies.
2-3 months research + 2-3 months web-portal build
RepRap Aftermarket Parts
3-18 monthsE3D built a >=GBP 10M business selling hotends to the RepRap ecosystem. Adjacent niches (precision belts, heaters, probes) are underserved.
3-5M USD revenue line in 3 years at 40-60% gross margin. Direct supply-chain intelligence — you become the supply chain.
GBP 500K-2M to stand up (machining/moulding, warehouse, ops)
Printed Electronics Investment
0-12 monthsNano Dimension, Voltera, and Optomec are closest to closing the electronics gap. Even a minor stake or advisory relationship gives priority access.
Asymmetric upside if the electronics gap closes. De-risking option regardless of timeline.
$100K-10M depending on vehicle
Humanoid Robot BoM Teardown
0-6 monthsThere is no good public BoM for any humanoid robot. Buy a Unitree G1, disassemble, and publish. Feeds directly into the 3D portal.
Media value (teardown videos draw traffic) + reference-data value (every future analysis cites the teardown).
~$20-30K for hardware + 2 months of teardown + publishing
Self-Replication Calculator
0-9 monthsA user-facing interactive tool that, given a machine's BoM, computes P%/M%/C% against our taxonomy. Integrates with HUMA's 3D humanoid portal.
Portal engagement, inbound founder/investor conversations, direct data on what machines people care about most.
1-2 months frontend + backend engineering
The Lineage
Self-replication has a 75-year intellectual history. These are the milestones that inform everything on this page.
RepRap
Adrian Bowyer's RepRap project created the world's first practical self-replicating manufacturing machine — a 3D printer that prints parts for copies of itself. The entire consumer 3D printing industry descends from RepRap.
von Neumann
John von Neumann formalised the kinematic self-reproducing automaton — proving that a machine could, in principle, build an exact copy of itself including the instructions for doing so. Published posthumously in 1966.
NASA Lunar Factory
NASA's Advanced Automation for Space Missions study (CP-2255) designed a self-replicating lunar factory that could build copies of itself from regolith. The most detailed engineering analysis of autotrophic machine replication ever conducted.
Open Source Ecology
Marcin Jakubowski's Global Village Construction Set — 50 industrial machines that can build a small, sustainable civilization with comfort. The machines can fabricate parts for each other, approaching distributed Class-5 replication.