Free Humanoid Platform — an open hardware, open firmware, open data humanoid robot reference design. The seventeenth morphology in the OpenLoco ecosystem. Shielded by the Free Humanoid Corpus.
A 1.6 m, ~50 kg humanoid expressed as a single OpenLoco UDD descriptor. The descriptor compiles to URDF, MJCF, STL meshes, BOM, and assembly instructions through the existing OpenLoco toolchain. Every non-trivial design choice cites a specific entry in the Free Humanoid Corpus by id, so that the choice is defensible against patent assertion as anticipated by prior art.
This document is the canonical full-system spec. Where the platform must choose between alternatives, this document maps the option space supported by the corpus rather than committing the architecture. Load-bearing engineering decisions that require user input are explicitly flagged as TBD (architectural call).
Open everything. Hardware, firmware, descriptors, BOM, control software, simulation models, training data, documentation. CERN-OHL-S 2.0 for hardware, Apache 2.0 for software, CC-BY-SA 4.0 for documentation, CC0 1.0 for descriptors and datasets.
The descriptor is the artifact, not the robot. A UDD JSON file is the source of truth. URDF, MJCF, STL, BOM, and assembly instructions all fall out of the descriptor through the OpenLoco compiler. Any platform that wants to fork ours starts by forking the descriptor.
Modularity at the interface, not at the implementation. Actuator, sensor, compute, and safety blocks are interface-defined. The reference implementations are the recommended path; alternatives are first-class. The HAL is the contract.
Prior-art-shielded design choices. Every non-trivial choice cites a corpus entry. If the corpus does not anticipate a choice, the choice is not made — either an entry is added to the corpus, or a different choice is made.
Reference design, not product. The deliverable is a buildable design that anyone can fork. Not a robot we sell. Not a service we provide. Not a brand we defend.
Fail safe to inert. Every subsystem has a defined fail-safe state. The default answer to a fault is “drop to a stable pose, latch brakes, kill high-voltage rails.” Safety supervisor (MathGround default) is non-bypassable from the policy layer.
Deterministic where it matters, probabilistic where it must. Low-level control, kinematics, balance, and safety are deterministic. Perception and policy are probabilistic. The boundary between them is sharp, not blended.
Human-scale, but not human-replacing. The platform is sized for human environments and tooling, not for humanoid uncanny-valley applications. The morphology is a means (compatibility with the built environment), not an end.
This is what already exists in open-source or public-domain prior art and what we build on. Each row cites the corpus entry id that establishes the prior art.
| Source (corpus id) | What it gives us |
|---|---|
wabot-1 (Waseda 1973) |
The original full-scale anthropomorphic biped. Full mechanism disclosure. 53-year prior art on humanoid bipedal kinematic topology. |
honda-e0 through honda-e6 (1986–1993) |
The Honda E-series chain. Static-walking-then-dynamic-walking biped lineage. 40-year prior art on the canonical humanoid leg kinematics (hip 3-DoF + knee + ankle 2-DoF). |
honda-p1, honda-p2, honda-p3 (1993–1997) |
First self-balancing autonomous full humanoids. ZMP-based dynamic balance disclosed publicly Dec 1996 (honda-p2). 28-year prior art that nukes most modern bipedal-locomotion claims. |
asimo (2000), hubo (2002) |
Mass-produced humanoid lineage with full kinematic disclosures. Publicly demoed; KAIST HUBO has academic publication chain. |
hrp-2 through hrp-5p (2002–2018) |
AIST/Kawada HRP series. Open academic disclosure of an industrial-scale humanoid through 2018. |
pal-talos (2017), nasa-valkyrie (2013), robonaut-2 (2011) |
Industrial / NASA full humanoids with substantial published technical disclosure. Robonaut 2 is on the ISS — disclosed in space. |
cassie-osu (2017), atrias (2013), digit-meta (2020) |
OSU-lineage spring-mass dynamic bipeds. Foundational for the modern walking-controller subspace. |
berkeley-humanoid (2024), k-scale-os (2024), unitree-h1 (2023) |
Recent academic / commodity humanoid disclosures. Berkeley Humanoid is open-source academic; Unitree H1 has full published spec. |
darwin-op (2010), poppy-humanoid (2012), inmoov (2012), reachy (2020) |
Open-source small/desk humanoid lineage. Permissive licenses, published BOMs, full CAD. |
Implication for the platform. The kinematic topology of a 1.6 m bipedal humanoid — head + torso + 2× 7-DoF arms + 2× 6-DoF legs + 2× anthropomorphic hands — is fully exhausted as patentable subspace. Any patent claim on humanoid kinematic topology that issues post-2025 is anticipated by at least one entry in the chain above.
This is the patent-thicket-densest subsystem and the single highest-leverage attack surface. The corpus carries the deepest chains here.
Cycloidal reducers (corpus chain anchored at sumitomo-cyclo 1937)
| Source (corpus id) | What it gives us |
|---|---|
sumitomo-cyclo (1937) |
The cycloidal reducer. 89-year prior art. Sumitomo Heavy Industries CYCLO drive, all original patents long expired. The canonical anchor. |
apptronik-apollo, sanctuary-phoenix, sanctuary-phoenix-gen6, tesla-optimus |
Modern humanoids using cycloidal in production. The corpus entries enumerate the patents asserted around these implementations. None of those patents survive 102/103 against sumitomo-cyclo. |
Recommendation. Cycloidal as the default for high-torque hip and knee joints. Direct from sumitomo-cyclo lineage; trivially defensible. TBD (architectural call): specific reducer ratio, sourcing (commodity Chinese cycloidals are now $80–$200 retail), and whether to design our own (full corpus shielding) or buy commodity (assembly speed).
Harmonic-drive (corpus chain anchored at honda-e0 1986)
| Source (corpus id) | What it gives us |
|---|---|
honda-e0 (1986) |
First disclosed harmonic-drive use in a humanoid. 40-year prior art. |
asimo, hubo, hrp-2..5p, robonaut-2, pal-talos, nasa-valkyrie |
Full chain through 2018 of harmonic-drive humanoids with published specs. |
Recommendation. Harmonic-drive as a secondary option for the shoulder and elbow joints where backlash matters and torque is moderate. Strain-wave geometry is patent-encumbered by Harmonic Drive Systems but the use of harmonic drives in humanoids is fully prior-arted.
Quasi-direct-drive (QDD) (corpus chain anchored at mit-cheetah-2 2014, cassie-osu 2017, mini-cheetah 2019)
| Source (corpus id) | What it gives us |
|---|---|
mit-cheetah (2009), mit-cheetah-2 (2014), mit-cheetah-3 (2017), mini-cheetah (2019) |
Foundational QDD architecture. 11-year prior art lineage with complete academic disclosure. |
cassie-osu (2017), digit-meta (2020), atrias (2013) |
QDD adaptation to bipeds. |
berkeley-humanoid (2024), k-scale-os (2024) |
Open-source QDD humanoid lineage. |
mjbots-moteus (2019), odrive (2017), simplefoc (2020) |
Open BLDC controllers. The unlock for cost. |
Recommendation. QDD as the default for the ankle (low-impedance, high-bandwidth ground reaction) and as an alternative to cycloidal for hip/knee where lower cost beats higher torque. Open BLDC controller (mjbots-moteus recommended for production-quality CAN-FD; odrive or simplefoc for hobbyist-tier) running FOC.
Tendon-driven / cable-driven (corpus chain anchored at shadow-dexterous-hand 2002, dlr-hand-ii 2001)
| Source (corpus id) | What it gives us |
|---|---|
shadow-dexterous-hand (2002), shadow-hand (2002) |
24-year prior art on tendon-routed anthropomorphic hand. Full mechanism disclosure. |
dlr-hand-ii (2001), dlr-justin (2009) |
DLR’s dexterous-hand and tendon-arm chain. |
pisa-iit-softhand (2012) |
Underactuated soft hand — synergy-based control. |
Recommendation. Tendon-driven for the hands and the wrist specifically, where cycloidal/harmonic-drive packaging is too bulky and the joint count is high. Tendon-driven is not recommended for the major leg or arm joints — too much friction, backlash, and routing complexity for the platform’s reliability target.
Actuator architectural call. TBD (architectural call): the platform’s default actuator distribution. The trade space:
sumitomo-cyclo); QDD at ankle (low impedance, shielded by mini-cheetah); harmonic-drive at shoulder and elbow (moderate torque, low backlash, shielded by honda-e0); tendon-driven at hands and wrist (shielded by shadow-dexterous-hand).The recommendation is Option C, but this is the call the user (David) should make. The descriptor is currently structured to support Option C and can be re-derived for A or B by editing the actuator_slots block.
| Source (corpus id) | What it gives us |
|---|---|
pomerleau-alvinn (1989) |
First end-to-end neural-network policy from camera to control. 35-year prior art on learned-from-pixels control. |
gelsight (2009), biotac-syntouch (2008), howe-cutkosky-tactile-1989 (1989) |
Full chain on tactile fingertip sensing. howe-cutkosky-tactile-1989 is a 36-year academic anchor. |
openai-rt-2 (2023), open-x-embodiment (2023) |
Foundational VLA (vision-language-action) prior art. Open arXiv / open dataset disclosure. |
act-aloha (2023), mobile-aloha (2024), diffusion-policy (2023) |
Imitation-learning-from-teleop foundational disclosures. |
openai-dactyl (2018) |
Sim-to-real RL with domain randomization on dexterous hand. |
Recommendation. Stereo cameras (commodity Luxonis/DepthAI or equivalent) + IMU at every link + force-torque at wrist + tactile (GelSight-style or capacitive) at fingertips. Perception runtime delegated to OpenLoco’s existing skill-graph runtime (see openloco’s v0.7.8 perception stack: SkillGraphRuntime, StubDetector, VilaQueryStub).
TBD (architectural call): depth modality. Stereo (passive, low-power, range-limited) vs. structured-light (better short-range, room-light-dependent) vs. ToF (medium range, narrow FoV). Stereo is the recommended default for cost and corpus-shielding.
| Source (corpus id) / OpenIE property | Role |
|---|---|
| Joule SOM (OpenIE recommended) | Default on-robot compute. Heterogeneous SoM with deterministic-LUT inference path. The descriptor’s compute block points at Joule SOM as the reference implementation. |
| Off-the-shelf alternatives (Jetson Orin, Intel NUC, Raspberry Pi 5 + Coral) | Swappable at the HAL boundary. |
The compute integration point is structurally clean: the firmware HAL exposes a small set of services (sensor reads, motor commands, safety hooks, perception result subscription) and the SoM choice is invisible above the HAL. Joule SOM is the recommended default; alternatives are first-class.
This is one of the highest-leverage subsystems and the corpus chain is the deepest single chain in the entire commons.
| Source (corpus id) | What it gives us |
|---|---|
asimov-positronic-robots (1940) |
86-year fictional anchor. Three Laws as inviolable hard constraints. The asimovs-zeroth-law (1985) extends. |
williamson-folded-hands (1947) |
Classic “safety constraint failure mode” prior art (the Humanoids over-protect). |
frankenstein (1818), rur-rossums-robots (1920) |
Earlier fictional anchors on autonomous-machine safety failure. |
hal-9000 (1968), data-tng (1987), robocop-1987 (1987) |
Mid-chain fictional anchors. RoboCop’s “Prime Directives” are an explicit Simplex-architecture-anticipating disclosure. |
sherman-simplex-architecture (1995) |
Foundational academic. The Simplex architecture: a verified safety controller as a fallback when the high-performance controller violates invariants. |
reachability-analysis-safe-control (2005), control-barrier-functions (2007) |
Formal-method safety controllers. CBFs are the modern standard. |
iso-10218-collaborative-robots (2006) |
Industrial standards anchor. |
runtime-assurance-rta (2010) |
AFRL runtime assurance pattern — bridges Simplex to flight-control. |
shielding-rl (2018) |
Shielded RL — formal safety on learned policy. |
Implication for the platform. Any patent on a “safety supervisor for physical AI” issued after 2010 faces an 80-year-deep chain of fictional anticipation plus a 30-year-deep chain of formal academic prior art. The architecture is fully shielded.
Recommendation. MathGround as the default safety supervisor implementation. It instantiates the sherman-simplex-architecture pattern with control-barrier-functions for the verified fallback controller, runs invariants on a separate compute domain from the high-performance controller, and is non-bypassable from the policy layer. Alternative: any open-source CBF implementation; the architectural pattern is what matters, not the implementation.
| Source (corpus id) | What it gives us |
|---|---|
mit-cheetah series, cassie-osu, digit-meta, atrias |
Open whole-body MPC for legged locomotion. |
act-aloha, mobile-aloha, diffusion-policy |
Imitation learning for manipulation. |
openai-rt-2, open-x-embodiment, physical-intelligence-pi-zero, skild-foundation-model, covariant-rfm |
VLA / foundation-model policies. |
openai-dactyl |
Sim-to-real RL on dexterous manipulation. |
Recommendation. Whole-body MPC (textbook formulation, no patentable claims) for low-level walking and balance, layered under an RL-trained policy for terrain-aware gait selection and locomotion mode switching. Manipulation policy: imitation-learning baseline (ACT or diffusion policy variant) with a VLA foundation-model option for higher-level task planning. The corpus shields all of this.
TBD (architectural call): policy architecture commitment. The trade space is between (a) classical MPC + scripted manipulation behaviors (interpretable, deterministic, low data requirement), (b) RL for locomotion + IL for manipulation (current academic best-practice), and (c) full VLA (highest ceiling, highest data and compute requirement). Recommendation is (b) with a path to (c).
| Source / OpenIE property | Role |
|---|---|
ROS 2 (corpus tag software-ros2) |
Onboard middleware, off-board orchestration. |
JANUS-equivalent local protocol (the corpus’s software-mjbots-stack analog) |
Robot-internal CAN-FD bus protocol for actuator and sensor traffic. |
| WebRTC / gRPC | Off-robot teleop and telemetry. |
Nothing patent-thicketed here — ROS 2 is permissive open source, CAN-FD is an ISO standard, WebRTC and gRPC are open. No special prior-art shielding needed.
| Source (corpus id) | What it gives us |
|---|---|
hyundai-boston-dynamics-spot, spot-fuel-cell |
Power architecture for legged platforms; Spot fuel-cell is an alternative-chemistry anchor. |
unitree-h1, unitree-g1, apptronik-apollo, tesla-optimus |
Modern humanoid power architectures (Li-ion, hot-swap battery packs). |
Recommendation. Hot-swap Li-ion pack (~1 kWh, 48 V nominal), commodity-grade. Hot-swap is the corpus-anchored design choice. Pack chemistry is commodity — no patent-thicket exposure.
TBD (architectural call): hot-swap mechanical interface. Every modern humanoid uses a different incompatible swap interface; standardizing one across the platform is a mild commons contribution.
The platform should degrade to recoverable / recyclable materials. Battery contained in IP-rated module. Magnets in motors are recoverable. PCBs are standard recyclable. Composite shells should prefer recyclable thermoplastics (PETG, PA12) over thermosets where stiffness allows.
This is not a corpus-shielded posture — it is a values-aligned design constraint inherited from the OpenIE family’s environmental posture (see clean_fish ARCHITECTURE §2.7).
unitree-h1 ~1.80 m, apptronik-apollo ~1.73 m, digit-meta ~1.60 m, pal-talos ~1.75 m, nasa-valkyrie ~1.88 m, all anchored back to the Honda E/P series and ASIMO 1.20–1.30 m, and back to WABOT-1 1.93 m at 1973).unitree-h1 5.4 m/s announced peak).Standard 33-DoF humanoid topology, decomposed as:
wabot-1’s 26-DoF torso lineage)pisa-iit-softhand, shadow-dexterous-hand, dlr-hand-ii)Total: 33 DoF in the recommended baseline. This is the standard humanoid topology, exhaustively prior-arted.
The full kinematic tree, with link masses, inertias, joint limits, and effort/velocity bounds, lives in descriptor/free-humanoid.udd.json. That descriptor is the source of truth; this section is summary.
Frame: aluminum 6061 / 7075 mix for high-load structural members; carbon fiber composite for shells. Detailed CAD: TBD — the descriptor encodes link inertias and origins; the detailed CAD is the next deliverable.
Corpus citations:
wabot-1, honda-p2, honda-p3 for foundational humanoid frame topology.unitree-h1, apptronik-apollo, berkeley-humanoid for modern lightweight frame implementations.See §1.2 for the option space and recommendation. The descriptor encodes a placeholder distribution consistent with Option C (hybrid). The actuator block is structured as actuator_slots per OpenLoco UDD convention, with each slot specifying:
{
"name": "...", // matches a joint name
"joint": "...",
"tier": 2, // OpenLoco tier: 0 Dynamixel, 1 SimpleFOC, 2 Moteus
"type": "cycloidal" | "harmonic_drive" | "qdd" | "tendon" | "linear",
"reducer_ratio": ...,
"peak_torque_nm": ...,
"continuous_torque_nm": ...,
"max_velocity_rad_s": ...,
"corpus_citation": "sumitomo-cyclo" | "honda-e0" | "mit-cheetah-2" | ...
}
sensing-proprioceptive-actuator.sensing-imu. Anchored at honda-p2 and the entire HRP/ASIMO chain.sensing-force-torque.howe-cutkosky-tactile-1989 → biotac-syntouch → gelsight.sensing-stereo-camera. Anchored at pomerleau-alvinn (1989) for camera-to-control.sherman-simplex-architecture). Perception runtime on its own compute domain (latency-decoupled from control).Default implementation: MathGround (OpenIE property) running the Simplex pattern.
The supervisor enforces a small set of formally specified invariants:
When any invariant is violated, the supervisor takes over, drives the robot to a defined safe pose (sitting, kneeling, or supported standing depending on context), latches the brakes, and reports the violation upstream. The high-level controller cannot override the supervisor. This is the sherman-simplex-architecture pattern with control-barrier-functions for the formal invariants.
Corpus citation chain (full): asimov-positronic-robots (1940) → williamson-folded-hands (1947) → hal-9000 (1968) → asimovs-zeroth-law (1985) → data-tng (1987) → robocop-1987 (1987) → sherman-simplex-architecture (1995) → reachability-analysis-safe-control (2005) → iso-10218-collaborative-robots (2006) → control-barrier-functions (2007) → runtime-assurance-rta (2010) → shielding-rl (2018).
Onboard perception pipeline:
foothold_map, ground_plane_fit, obstacle_height skills are the reference implementations.SkillGraphRuntime (OpenLoco’s skill graph). Corpus: openai-rt-2, open-x-embodiment, physical-intelligence-pi-zero.gelsight.pisa-iit-softhand (synergy-based), shadow-dexterous-hand (full-DoF reference), dlr-hand-ii.See §1.6. Layered:
gait_harness (the same harness that runs the ANYmal-WL and Mini Whegs gaits). Corpus: mit-cheetah-2, cassie-osu, mini-cheetah, digit-meta.act-aloha, mobile-aloha, diffusion-policy. Optional VLA upgrade: openai-rt-2, physical-intelligence-pi-zero, skild-foundation-model.Hot-swap Li-ion pack, ~1 kWh nominal, 48 V. Onboard 24 V and 12 V DC-DC rails for compute and sensors. Estimated runtime 1.5–3 hours depending on workload.
Hot-swap interface: committed (§9 #6) to a CC0 commons specification designed by this project and contributed upstream to OpenLoco UDD as an extension proposal. The spec describes mechanical mount, electrical contacts, BMS-handshake protocol, and thermal interface. Modeled on the shoal cartridge interface pattern. Avoids commercial battery-pack vendor lock-in.
Batteries IP-rated and removable for recycling. Magnets in motors recoverable. Composite shells preferring thermoplastics over thermosets. PCBs standard-recyclable. Documentation includes a takedown / disassembly manual matching the assembly manual.
The single source of truth is descriptor/free-humanoid.udd.json. It is OpenLoco UDD-compliant. The OpenLoco compiler can already turn it into:
free-humanoid.urdf — ROS 2 / RViz / Gazebofree-humanoid.xml — MJCF for MuJoCo / MJXbom.csv — bill of materialsassembly.md — assembly instructionsTo compile:
# from an OpenLoco checkout
cargo run -- validate /path/to/free-humanoid.udd.json
cargo run -- generate /path/to/free-humanoid.udd.json --all --output-dir out/
cargo run -- bake /path/to/free-humanoid.udd.json --output-dir out/meshes/
The descriptor currently uses placeholder mass/inertia values structured to OpenLoco UDD conventions. Iterating these to physical-build accuracy is a load-bearing follow-up task, gated on the actuator-distribution and mass-budget architectural calls.
free-humanoid-platform/
README.md
ARCHITECTURE.md (this document)
CONTRIBUTING.md
LICENSE-HARDWARE CERN-OHL-S 2.0
LICENSE-SOFTWARE Apache 2.0
LICENSE-DOCS CC-BY-SA 4.0
LICENSE-DATA CC0 1.0
descriptor/
free-humanoid.udd.json the source of truth
README.md
chassis/ mechanical CAD pointers, hull, kinematics
actuators/ actuator family options and specs
electronics/ PCB pointers, power tree, Joule SOM integration
firmware/ HAL, behaviors, safety supervisor (MathGround), comms
control/ whole-body MPC, RL policy weights, manipulation policy
sim/ MuJoCo / Stonefish / Drake model pointers
docs/ additional design notes
datasets/ pointers to training data sources
prior-art/
INDEX.md corpus entries this platform depends on, by subsystem
Not covered by existing prior art and requiring genuine new disclosure:
The UDD schema extension for humanoid-specific actuator slots. The existing UDD schema handles tier-2 Moteus QDD; humanoid-specific extensions for cycloidal and harmonic-drive actuator types need to be added (and contributed back upstream to OpenLoco).
The hand descriptor. No existing OpenLoco morphology has a 6-DoF underactuated 5-finger hand. The synergy-reduction representation in UDD is novel and needs to be designed in collaboration with OpenLoco maintainers.
The MathGround integration shim. MathGround as a Simplex supervisor for a UDD-described robot is a new integration. The interface contract is defined here; the implementation is downstream.
The hot-swap battery interface mechanical standard. Optional small commons contribution; the interface is novel only in that it is standardized.
Everything else is fork, port, or integration of existing open work, with corpus citations.
gait_harness for closed-loop walking rollouts.pisa-iit-softhand, shadow-dexterous-hand. Build the hand first — it is the highest-cost-per-DoF subsystem and the highest-leverage demonstration that the open posture can match commercial closed designs.gelsight corpus citation).clean_fish ARCHITECTURE §9.Three reasons, the same three the clean_fish architecture cites and which the OpenIE family applies to every substrate.
Trust. Embodied physical AI in human environments requires radical transparency. Anyone deploying a humanoid in a workspace must be able to inspect the safety supervisor, the control policy, the perception stack, and the actuator firmware. Closed-source physical AI is a trust failure waiting to happen.
Distribution. Humanoids will not be solved by a few well-capitalized incumbents. The space is too broad — industrial, domestic, hazardous-environment, eldercare, agricultural, construction. An open kit replicated by hundreds of universities, labs, and small companies covers the space; a half-dozen closed silos do not.
Compounding. Every fork makes the next fork easier. Every corpus citation makes the next platform’s defense easier. Every UDD descriptor strengthens OpenLoco. The commons compounds; the silos do not.
This is the OpenIE thesis applied to humanoids. The descriptor compiler, the prior-art commons, the safety supervisor, the deterministic compute spine — they all converge on a humanoid that anyone can build, anyone can fork, and no one can enclose.
The ten load-bearing decisions, resolved. Each commitment is shielded by the cited prior art chain so that adopting it carries minimal patent-thicket risk. Subsystem sections above are written to these commitments.
| # | Decision | Commitment | Shielding chain (corpus ids) |
|---|---|---|---|
| 1 | Actuator distribution | Hybrid. Cycloidal at hip + knee; harmonic-drive at shoulder + elbow; QDD at ankle; tendon at wrist + hand. | sumitomo-cyclo (1937, 89 yr) for hip/knee; honda-e0 (1986, 40 yr) and dlr-toro (2014) for shoulder/elbow; mit-cheetah-2 (2014), mini-cheetah (2019), cassie-osu for ankle QDD; da-vinci-knight (1495, 528 yr) plus shadow-dexterous-hand (2002), dlr-hand-ii (2001) for wrist + hand tendon. The hybrid lets each joint be shielded by its deepest available prior-art chain rather than forcing a one-size choice. |
| 2 | Mass budget | 50 kg reference build. 30 kg “academic-light” and 80 kg “industrial-heavy” published as descriptor variants. | wabot-1 (1973), Honda E/P chain (honda-e0–honda-p3 1986–1997), asimo, HRP series (hrp-2 2003 → hrp-5p 2018), nasa-valkyrie, pal-talos, cassie-osu, digit-meta, berkeley-humanoid. The 50 kg target hits the median of the academic-platform mass distribution; nothing about the choice is novel. |
| 3 | Default policy stack | Classical MPC + RL locomotion + IL manipulation. Layered as in §2.10. Full VLA documented as a recommended upgrade path with its own shielding. | MPC: khatib-operational-space (1987), featherstone-rdf (1987), kajita-lipm (2001). RL locomotion: cassie-osu, mini-cheetah, hwangbo-anymal-sim2real (2019). IL manipulation: act-aloha (2023), mobile-aloha (2024), diffusion-policy (2023). Optional VLA: openai-rt-2 (2023), open-x-embodiment (2023), openvla (2024), physical-intelligence-pi-zero. |
| 4 | Default end-effector | Underactuated 5-finger hand (synergy-reduced). Parallel-jaw and full-DoF Shadow-class hands published as alternative descriptor variants. | da-vinci-knight (1495), pisa-iit-softhand (2014, synergy reference), shadow-dexterous-hand (2002), dlr-hand-ii (2001). Synergy-driven underactuation has the simplest BOM and the strongest academic anchor for an open reference. |
| 5 | Depth modality | Stereo cameras (head-mounted, paired with on-fingertip GelSight tactile). | pomerleau-alvinn (1989) for camera-to-action stereo; howe-cutkosky-tactile-1989, biotac-syntouch (2008), gelsight (2009) for fingertip tactile. Structured-light and ToF carry heavier patent thickets (commercial entanglements with Apple, Intel, Microsoft, Sony); stereo + GelSight is the demonstrably-clean depth+contact stack. |
| 6 | Hot-swap battery interface | Design our own commons spec, contribute upstream to OpenLoco UDD. Modeled on the shoal cartridge interface pattern. CC0. | Commercial battery-pack standards carry vendor-IP entanglements; designing the spec ourselves and donating it to a CC0 commons is the anti-patent-thicket move and matches the OpenIE family pattern (shoal’s gut-cartridge interface, openloco’s UDD descriptor format). The spec will be a short OpenLoco extension proposal alongside the actuator-slot extensions. |
| 7 | Cycloidal sourcing | Commodity buy for v0 (Sumitomo CYCLO product line, Onvio, Spinea). Design-our-own held as a future track once manufacturing volume justifies it. | sumitomo-cyclo (1937) plus the academic/industrial chain since means commodity sourcing is bulletproof; engineering effort is better spent on integration and the higher-leverage hand and safety supervisor work. |
| 8 | Hand variant for first physical build | Underactuated 5-finger matching the v0 default (commitment 4). Shadow-class held as a research-grade upgrade path for Phase 3+. | pisa-iit-softhand (2014). Lower BOM, simpler training data, strong academic prior art. |
| 9 | Reference-design vs specific-build governance | Reference design. The descriptor (free-humanoid.udd.json, CC0) is the canonical artifact. Specific physical builds — with chosen vendor parts, BOM-specific dimensions, etc. — are downstream tenants. |
Matches OpenLoco’s substrate-vs-tenant pattern: OpenLoco is the descriptor compiler, individual robots (and now this platform) are tenants. The platform repo ships the reference; vendor-specific builds live in tenant repos that pin their BOM. |
| 10 | Joule SOM commitment depth | Joule SOM as recommended-default within an abstract HAL. The descriptor’s compute block specifies the abstract HAL interface; Joule SOM is one implementation. Jetson, Coral, NUC, Pi 5 are documented alternative implementations. |
The HAL abstraction is the OpenLoco-substrate pattern applied to compute. The platform isn’t locked to Joule, but Joule is the recommended path because it’s the only implementation today that gives deterministic-LUT inference + RISC-V housekeeping + safety co-processor in a single module. |
These commitments are this document’s defaults. Subsystem sections (§2) reflect them in detail. Future amendments require a corpus-citation update demonstrating that any new shielding chain is at least as deep as the chain it replaces.
Free Humanoid Platform — the seventeenth OpenLoco morphology — scaffold v0.1 — 2026-05-06.